AI-based intelligent response method for IoT traffic demand peaks

Abstract

The present invention discloses an AI-based intelligent response method for IoT traffic demand peaks, which relates to the field of IoT traffic management. The method comprises collecting historical usage data and related business information of multiple traffic pools to form multi-dimensional feature information, and inputting the information into an AI traffic prediction model to obtain traffic demand prediction results of each traffic pool; calculating the predicted utilization rate based on the current used traffic and the prediction results, triggering a replenishment operation and generating a cross-pool replenishment plan when the utilization rate reaches a preset threshold, and executing the traffic pool replenishment operation according to the replenishment instruction, thereby solving the problems of untimely replenishment during traffic peak periods and low scheduling efficiency.

Description

Technical Field

[0001] The present invention relates to the field of Internet of Things traffic management, and more specifically, to an AI-based intelligent response method for Internet of Things traffic demand peaks. Background Art

[0002] With the rapid growth of IoT devices, various sensors, smart terminals, and data collection nodes are being widely deployed in areas such as transportation, smart cities, and industrial control. The resulting network traffic exhibits complex characteristics involving multiple services, multiple scenarios, and multiple time series. During major events or when sudden services are launched online, traffic demand often rises sharply in a short period of time, posing a significant challenge to capacity planning and resource scheduling for traffic pools. This type of peak traffic is characterized by wide geographical distribution and diverse service types. Furthermore, because prediction models struggle to accurately capture sudden changes, traditional static reservations or experience-based replenishment strategies struggle to meet demand in a timely manner. Insufficient capacity can lead to service quality degradation or interruptions, while frequent replenishment can waste resources and unmanageable costs. These issues urgently require multi-dimensional feature analysis and intelligent prediction to improve the efficiency of proactively predicting and responding to traffic peaks. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide an AI-based intelligent response method for IoT traffic demand peaks to solve the problems mentioned in the background technology.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions:

[0005] An AI-based intelligent response method for IoT traffic demand peaks, including:

[0006] Collect historical usage data and business information from multiple traffic pools to form multi-dimensional feature information;

[0007] Inputting the multi-dimensional feature information into an AI traffic prediction model to obtain traffic demand prediction results for the multiple traffic pools respectively;

[0008] Based on the current used traffic and traffic demand forecast results of each traffic pool, calculate the predicted utilization rate of each traffic pool;

[0009] When the traffic pool utilization reaches the preset threshold T When the traffic pool needs to be replenished, the replenishment operation of the traffic pool is triggered, and a cross-pool replenishment plan is generated, which includes the allocation of replenishment packages for each traffic pool that needs to replenish traffic;

[0010] The replenishment package distribution in the cross-pool replenishment plan is sent as a replenishment instruction to perform a traffic pool replenishment operation.

[0011] In some embodiments, the multi-dimensional feature information includes time features, business features, customer features, scenario features, and traffic pool features, and the traffic pool features include the unit price of the supplementary package, the historical supplementary delay rate, and the remaining budget.

[0012] In some embodiments, the AI ​​traffic prediction model is a multi-layer long short-term memory network model, which is used to generate a prediction sequence for each traffic pool. , calculated using the following formula:

[0013] ;

[0014] Where i is the traffic pool number, t is the current time point, k is the prediction step, and n is the number of feature dimensions. is the p-th dimension eigenvalue, is the corresponding weight.

[0015] In some embodiments, the predicted utilization of each traffic pool Calculated using the following formula:

[0016] ;

[0017] Where, is the used flow of the i-th flow pool before time t, m is the number of prediction time steps, is the total capacity of the No. i flow pool, j is the index of the prediction time step.

[0018] In some embodiments, the method uses a convex optimization algorithm to generate a cross-pool replenishment plan by solving the following linear programming problem:

[0019] ;

[0020] Where, is the replenishment amount allocated to the i-th flow pool, is the unit price of the No. i traffic pool supplement package, To supplement the global budget, N is the total number of traffic pools.

[0021] In some embodiments, the AI ​​traffic prediction model is trained by the following steps:

[0022] Using historical traffic data and corresponding multi-dimensional feature information, we minimize the mean square error between the predicted value and the actual value as the training goal and optimize the corresponding weights. .

[0023] In some embodiments, the preset threshold T Artificial setting 。

[0024] In some embodiments, the preset threshold T The following method is used for setting: using reinforcement learning to take the predicted utilization rate, replenishment cost and breach penalty of the traffic pool as state input, and autonomously output the preset threshold used for the global optimization scheduling.

[0025] In some embodiments, the design and training process of the reinforcement learning model includes:

[0026] The state is defined as a state vector containing the predicted utilization of each traffic pool, the unit price of the supplementary package, and the penalty for breach of contract;

[0027] Actions are defined as adjusting preset thresholds T The value of

[0028] The reward function is designed to be the negative of the weighted sum of the supplementary cost and the default penalty;

[0029] The training process is carried out in a simulated environment, and the threshold strategy is optimized through multiple iterations to minimize the long-term cumulative cost.

[0030] In some embodiments, after the replenishment operation is performed, the actual effect of the replenishment is recorded, including the change in utilization rate after the replenishment; and the recorded replenishment effect data is used to provide feedback information for subsequent traffic prediction and replenishment strategy optimization.

[0031] The advantages of the present invention over existing technologies are that it first extracts multi-dimensional features from the historical usage data and related business information of multiple traffic pools, and inputs them into a multi-layer long short-term memory network model with an attention mechanism to generate a traffic demand forecast sequence for each traffic pool. Based on the current used traffic and the forecast results, the predicted utilization rate is calculated. When the utilization rate reaches a preset threshold, the system automatically triggers traffic pool replenishment and generates a cross-pool replenishment plan. After the replenishment instruction is issued, it is immediately executed, significantly solving the problems of insufficient capacity and delayed replenishment during peak traffic periods, effectively improving the real-time and accuracy of scheduling, and avoiding service interruptions or over-replenishment caused by forecast deviations. Furthermore, combined with an auxiliary mechanism, a convex optimization algorithm is used to achieve the optimal allocation of cross-pool replenishment under global budget and cost constraints. Reinforcement learning is used to dynamically adjust the trigger threshold, enabling the system to autonomously optimize the response strategy based on the replenishment cost and penalty. After the replenishment operation, the utilization rate changes are recorded in real time and fed back to the prediction and scheduling model, forming a closed-loop optimization of prediction and response, continuously enhancing the system's adaptability and economic efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is the overall flow chart of the present invention;

[0033] Figure 2 It is a detailed flow chart of data collection of the present invention;

[0034] Figure 3 It is a flow chart of the model prediction of the present invention;

[0035] Figure 4 It is a flow chart of the supplementary triggering and solution generation of the present invention. DETAILED DESCRIPTION

[0036] The specific embodiments of the present invention will be described below with reference to the accompanying drawings.

[0037] As shown in Figure 1, this invention provides an AI-based intelligent response method for IoT traffic demand peaks. This method aims to efficiently address sudden spikes in traffic demand within IoT systems through intelligent traffic prediction and resource scheduling, ensuring the rational allocation of traffic resources and the stability of system operations. The following is a detailed description of the invention's implementation process, covering multiple steps including data collection, model prediction, utilization calculation, supplementary triggering, solution generation, and instruction execution, with detailed explanations of each step.

[0038] In IoT systems, a traffic pool is a unit of shared traffic resources allocated to multiple devices or users. Each traffic pool has a certain total capacity, and devices gradually consume this capacity as they use network services. To accurately predict traffic demand, it's first necessary to comprehensively collect historical usage data and relevant business information from multiple traffic pools. This data is processed to form multi-dimensional feature information, which serves as the basis for subsequent analysis.

[0039] As shown in Figure 2, the collected data specifically includes several key aspects. Temporal characteristics reflect the periodicity and regularity of traffic usage, such as date, specific time (hour, minute), day of the week, and whether it is a holiday. This information can help capture seasonal or cyclical changes in traffic demand. Service characteristics involve specific attributes of the service, such as service type (video streaming, data transmission, etc.), service volume, and service priority. Traffic demand patterns for different services can vary significantly. Customer characteristics focus on the characteristics of the customer groups using the traffic pool, such as customer type (individual user or enterprise user), customer scale (number of devices), and customer activity (frequency of use). These factors influence the speed and scale of traffic consumption. Scenario characteristics relate to the specific context of traffic usage, including geographic location (urban or rural), network environment (4G or 5G), and device type (sensor or smart terminal). These conditions can have a direct impact on traffic demand. Furthermore, traffic pool characteristics are parameters closely related to traffic pool operations and management, including the unit price of replenishment packages (the unit cost of each replenishment), historical replenishment latency (the percentage of response time for replenishment operations), and remaining budget (the remaining funds available for replenishment). This information provides important information for resource allocation.

[0040] Through the systematic collection and preprocessing of the above data, such as normalization, missing value filling, or outlier removal, structured multi-dimensional feature information is ultimately formed. This feature information is represented in the form of vectors and provides input data for subsequent AI model predictions.

[0041] As shown in Figure 3, to predict future traffic demand for each traffic pool, this method employs an advanced AI traffic forecasting model, specifically a multi-layer long short-term memory (LSTM) network with an attention mechanism. The LSTM model is particularly well-suited for processing time series data and can effectively capture long-term dependencies in traffic usage patterns. The attention mechanism dynamically adjusts feature weights to enhance the model's focus on key information, thereby improving forecast accuracy.

[0042] For each flow pool i , the model generates its future k The traffic demand forecast sequence of time steps is recorded as The calculation process can be expressed by the following formula:

[0043] ;

[0044] In this formula, i represents the number of the traffic pool, t is the current time point, k is the prediction step (for example, predicting demand for the next hour or day), and n is the total number of feature dimensions. is the p-th dimension feature in time t The value of is the weight of the corresponding feature, which is automatically learned by the attention mechanism based on historical data. The core idea of ​​this formula is to perform weighted summation of multi-dimensional feature information to generate a future traffic demand forecast. The introduction of enables the model to dynamically identify which features (such as holidays or business types) have a greater impact on traffic demand, thereby improving the pertinence and accuracy of predictions.

[0045] In practical applications, the architecture of the model contains multiple LSTM layers, each layer is responsible for extracting features of different time spans, and finally integrates the information through the attention layer to output a prediction sequence. The model contains an input layer, which is responsible for receiving multi-dimensional feature information. These features may include time, business type, customer attributes, scenarios, and traffic pool-related data. The input layer organizes this information and passes it to subsequent layers to provide the model with the basic data required for prediction; the middle multi-layer LSTM layer is composed of multiple stacked LSTM units. LSTM is an improved recurrent neural network (RNN) that can effectively memorize long-term information and process time series data through gating mechanisms such as forget gates, input gates, and output gates. The multi-layer structure enables the model to learn more complex patterns and features. The number of layers is usually between 2 and 4, depending on the data complexity and performance requirements. The attention mechanism layer dynamically adjusts the weights of the features to focus the model on the features that have the greatest impact on the prediction results. In specific implementation, the attention mechanism may calculate the weight of each feature through a neural network (such as the formula ), thereby enhancing the targetedness and accuracy of the prediction. The model generates the final traffic demand forecast results at the output layer. Furthermore, to improve generalization, the model may also include a dropout layer (optional), which prevents overfitting by randomly dropping some neurons during training.

[0046] During the training process, historical traffic data and corresponding feature information are used to optimize the model parameters by minimizing the error (e.g., mean square error) between the predicted value and the actual value. During the training process, the attention mechanism gradually adjusts the feature according to its contribution to the prediction result. These weights usually range from [0, 1] and are normalized to ensure that the sum is 1.

[0047] Therefore, the prediction process can be summarized as follows:

[0048] 1. The system first extracts a set of key features (such as historical traffic values, service type codes, time tags, etc.) from each traffic pool at time t.

[0049] 2. The feature sequences of several past moments are sequentially input into the LSTM network. The LSTM automatically learns how these features change over time and, in the final step, outputs a "hidden state" vector representing the temporal context of the current traffic trend.

[0050] 3. Use a simple fully connected layer to map this hidden state into multiple sets of weight vectors , each set of weights corresponds to the importance distribution of each feature at a certain moment in the future (such as t+1, t+2...t+k).

[0051] 4. Multiply each weight vector by the feature vector at the current moment and add them together to obtain the traffic demand forecast value at the corresponding future moment.

[0052] 5. If you need to predict multiple steps, generate the weight vectors for all future steps at once and calculate them one by one. You can also use a rolling method to feed the predicted value back to the next step and repeat the above process.

[0053] In this way, LSTM is responsible for understanding "what features and how they change in time" will most affect future traffic. The fully connected layer then dynamically gives the feature weights at different prediction moments, and finally uses weighted summation to obtain the specific traffic prediction for each step.

[0054] As shown in Figure 4, after obtaining the traffic demand forecast results for each traffic pool, it is necessary to calculate the predicted utilization of each traffic pool in combination with the current used traffic to evaluate its future resource usage. The calculation formula is as follows:

[0055] ;

[0056] in, Indicates that as of the current time t, i The traffic that has been consumed by the traffic pool; m is the number of time steps in the forecast, i.e. the forecast horizon considered in the future; It is i The traffic pool will be in the future j Forecast traffic demand for time steps; It is i The total capacity of the traffic pool. This formula calculates the current used traffic plus the future m The percentage of the sum of the predicted demands in the time steps to the total capacity reflects the resource shortage of the traffic pool in the future.

[0057] The reason for this design is that it can take into account the current traffic usage status and future potential demand, providing a forward-looking utilization indicator. For example, if a traffic pool currently has a high amount of used traffic and the demand is expected to continue to increase in the future, then its predicted utilization may quickly approach or exceed the total capacity, indicating that intervention measures are needed. Parameters m The value of needs to be flexibly adjusted according to the business scenario. For example, for systems with high real-time requirements, m It can be set to a smaller value (such as the number of predicted steps within 1 hour), and for long-term planning, m It can be extended appropriately (such as 1 day or 1 week).

[0058] When the predicted utilization rate of a traffic pool Reaching or exceeding a preset threshold T When the system triggers the replenishment operation of the traffic pool. T is a key parameter used to balance resource utilization efficiency and replenishment cost. T If the setting is too low, it may lead to frequent replenishment and increase operating costs; T If it is too high, there may be a risk of service interruption due to insufficient traffic.

[0059] In order to intelligently determine the optimal threshold T , this method introduces reinforcement learning technology. The reinforcement learning model takes the predicted utilization rate, replenishment cost and default penalty of the traffic pool as state input, and gradually optimizes the output threshold required for global scheduling through interactive learning with the environment. Specifically, the model defines the current state as a vector, including the , Supplementary pack unit price and potential penalties for breach of contract (e.g., compensation for service interruption) that may result from insufficient traffic. The model's action is to adjust T The goal is to find a threshold strategy with the lowest long-term cumulative cost through trial and error. The reward function can be designed as the negative of the weighted sum of the supplementary cost and the default penalty, with the weights determined by business priorities.

[0060] During the training process, the reinforcement learning model simulates different traffic demand scenarios and tries different T value, and updates the strategy based on the effect of the supplementary operation (e.g. whether insufficient flow is avoided). T Based on experience, this threshold can be set to 80% or 90%. After multiple iterations, the model will converge to a dynamic threshold that adapts to specific business needs, typically fluctuating between 70% and 95%, depending on the capacity and usage patterns of the traffic pool. Of course, in other embodiments, this threshold can also be manually preset.

[0061] Once a replenishment operation is triggered, the system needs to generate a cross-pool replenishment plan to determine how to allocate traffic from global resources to meet demand. This method uses a convex optimization algorithm to achieve this goal by solving a linear programming problem. The optimization goal is to minimize the replenishment cost, which can be expressed mathematically as:

[0062] ;

[0063] The following constraints must also be met:

[0064] ;as well as:

[0065] ;

[0066] in, is assigned toi The amount of supplementary flow in the flow pool, is the unit price of the traffic pool supplement package, It is the global supplementary budget. N is the total number of flow pools. It should be noted that the constraints here assume that the supplementary flow Will increase the total capacity of the flow pool , thereby reducing the predicted utilization below the threshold T .

[0067] The significance of this linear programming problem is that, within a limited budget Allocate additional traffic at the lowest cost within the period, ensuring that all traffic pools will be m The utilization rate in each time step is controlled within a safe range. Since the objective function and constraints are linear, mature optimization tools (such as the simplex method or the interior point method) can be used to efficiently solve and obtain the optimal replenishment amount. .

[0068] After generating a cross-pool replenishment plan, the system converts the allocation of replenishment packages into specific replenishment instructions and sends them to the corresponding traffic pools for execution. These instructions may include directly purchasing new traffic packages and allocating them to the target traffic pools, or in some scenarios, transferring traffic from underutilized traffic pools to the ones in need of replenishment. During execution, the system records the actual results of the replenishment, such as changes in utilization after replenishment, to provide feedback for subsequent forecasting and optimization.

[0069] To ensure the effective implementation of the method, some key parameters need to be reasonably set and the AI ​​model needs to be fully trained. Usually k can be set to 1 hour to 7 days, and m can be a multiple of k. For example, if k is 1 day, then m can be 3 to 7, depending on the business's real-time and forward-looking requirements. The feature dimension n is determined by the collected data and usually includes time, business, customer, scenario, and traffic pool features, and the total number may be between dozens and hundreds. Determined by market price, for example, 10 yuan / GB to 20 yuan / GB, and the global supplementary budget It is set by the financial situation of the enterprise, such as 500GB or 1000GB per month.

[0070] Training AI traffic prediction models requires extensive historical data, with the goal of minimizing prediction error. After training, the model continuously updates its parameters through online learning to adapt to changing traffic patterns. Training reinforcement learning models requires building a simulation environment to simulate the state changes of the traffic pool and the execution of supplementary operations, optimizing threshold strategies through multiple iterations.

[0071] To illustrate the implementation of the method more intuitively, the following is a simplified application example. Assume that the system manages three traffic pools, the current time t=0, and the parameters of each traffic pool are as follows:

[0072] Traffic pool 1: used traffic 100GB, total capacity 200GB, refill package price 10 yuan / GB

[0073] Traffic pool 2: used traffic 150GB, total capacity 300GB, supplementary package price 8 yuan / GB

[0074] Traffic pool 3: used traffic 200GB, total capacity 500GB, supplementary package price 12 yuan / GB

[0075] Assuming the prediction step size k=1 and the number of time steps m=1, the AI ​​model predicts the traffic demand in the next time step as:

[0076] Traffic pool 1: 50GB

[0077] Traffic pool 2: 100GB

[0078] Traffic pool 3: 150GB

[0079] According to the predicted utilization formula, the calculation results are as follows:

[0080] Traffic pool 1: (100 + 50) / 200 × 100% = 75%

[0081] Traffic pool 2: (150 + 100) / 300 × 100% = 83.33%

[0082] Traffic pool 3: (200 + 150) / 500 × 100% = 70%

[0083] Assuming the threshold T = 80%, the predicted utilization of traffic pool 2 is 83.33%, which exceeds the threshold and needs to be replenished. =500GB, solve the linear programming problem with the following constraints:

[0084] ;

[0085] Calculated:

[0086] ;

[0087] Therefore, at least 12.5 GB needs to be added to traffic pool 2. Assuming 15 GB is added, the cost is 15 × 8 = 120 yuan. At this point, the total capacity of traffic pool 2 is 315 GB, and the predicted utilization rate is 250 / 315 × 100% ≈ 79.37%, which is lower than 80% and meets the requirement. The utilization rates of other traffic pools are already below the threshold, so no additional capacity is required.

[0088] In this way, the system can achieve optimal allocation of traffic resources at a lower cost and ensure service stability.

[0089] This invention combines AI prediction, reinforcement learning, and convex optimization techniques to achieve intelligent response to IoT traffic demands. Future optimizations of feature selection, model architecture, or supplementary strategies can be applied to improve prediction accuracy and resource utilization efficiency, based on actual application scenarios. For example, more real-time data could be introduced, or more complex deep learning models could be employed to address complex traffic patterns.

[0090] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. An AI-based intelligent response method for IoT traffic demand peaks, characterized by: include: Collect historical usage data and business information from multiple traffic pools to form multi-dimensional feature information; Inputting the multi-dimensional feature information into an AI traffic prediction model to obtain traffic demand prediction results for the multiple traffic pools respectively; Based on the current used traffic and traffic demand forecast results of each traffic pool, calculate the predicted utilization rate of each traffic pool; When the traffic pool utilization reaches the preset threshold T When the traffic pool needs to be replenished, the replenishment operation of the traffic pool is triggered, and a cross-pool replenishment plan is generated, which includes the allocation of replenishment packages for each traffic pool that needs to replenish traffic; Sending the replenishment package allocation in the cross-pool replenishment plan as a replenishment instruction to perform a traffic pool replenishment operation; The AI ​​traffic prediction model is a multi-layer long short-term memory network model used to generate a prediction sequence for each traffic pool. , calculated using the following formula: ； Where i is the traffic pool number, t is the current time point, k is the prediction step, and n is the number of feature dimensions. is the p-th dimension eigenvalue, is the corresponding weight; Predicted utilization of each traffic pool Calculated using the following formula: ； Where, is the used flow of the i-th flow pool before time t, m is the number of prediction time steps, is the total capacity of the No. i flow pool, j is the index of the prediction time step; A convex optimization algorithm is used to generate a cross-pool replenishment plan by solving the following linear programming problem: ； Where, is the replenishment amount allocated to the i-th flow pool, is the unit price of the No. i traffic pool supplement package, To supplement the global budget, N is the total number of traffic pools.

2. The AI-based IoT traffic demand peak intelligent response method according to claim 1 is characterized in that: The multi-dimensional feature information includes time features, business features, customer features, scenario features and traffic pool features. The traffic pool features include the unit price of the supplementary package, the historical supplementary delay rate and the remaining budget.

3. The AI-based IoT traffic demand peak intelligent response method according to claim 1 is characterized in that: The AI ​​traffic prediction model is trained through the following steps: Using historical traffic data and corresponding multi-dimensional feature information, we minimize the mean square error between the predicted value and the actual value as the training goal and optimize the corresponding weights. .

4. The AI-based IoT traffic demand peak intelligent response method according to claim 1 is characterized in that: Preset threshold T Artificial setting 。 5. The AI-based IoT traffic demand peak intelligent response method according to claim 1 is characterized in that: Preset threshold T The following method is used to set the state: using reinforcement learning to take the predicted utilization rate of the traffic pool, the replenishment cost and the penalty for breach of contract as the state input, and autonomously output the preset threshold T .

6. The AI-based IoT traffic demand peak intelligent response method according to claim 5 is characterized in that: The design and training process of the reinforcement learning model includes: The state is defined as a state vector containing the predicted utilization of each traffic pool, the unit price of the supplementary package, and the penalty for breach of contract; Actions are defined as adjusting preset thresholds T The value of The reward function is designed to be the negative of the weighted sum of the supplementary cost and the default penalty; The training process is carried out in a simulated environment, and the threshold strategy is optimized through multiple iterations to minimize the long-term cumulative cost.

7. The AI-based IoT traffic demand peak intelligent response method according to claim 1 is characterized in that: After performing the replenishment operation, the actual effect of the replenishment is recorded, including the change in utilization after the replenishment; the recorded replenishment effect data is used to provide feedback information for subsequent traffic prediction and replenishment strategy optimization.

Images