Artificial intelligence-based resource scheduling methods, devices, equipment, and storage media

Abstract

This invention provides a resource scheduling method, apparatus, device, and storage medium based on artificial intelligence, relating to the field of communication technology. The method includes: dividing a cell coverage area into multiple grids; employing an LSTM-based grid load prediction model to output the grid load prediction value at the start of the next scheduling cycle; allocating the total cell bandwidth resources to each grid proportionally based on the proportion of each grid's grid priority score to the sum of all grid priority scores, thus obtaining the schedulable resource amount for each grid in the current scheduling cycle; and for each grid, proportionally allocating the schedulable resources of that grid to each user in the current scheduling cycle based on the proportion of each user's weight within that grid to the sum of all user weights within that grid. The technical solution provided by this invention can dynamically sense changes in local load distribution and edge user demand, and adaptively adjust resource allocation.

Description

Technical Field

[0001] This invention relates to the field of communication technology, and in particular to an artificial intelligence-based resource scheduling method, an artificial intelligence-based resource scheduling device, a computer device, and a computer-readable storage medium. Background Technology

[0002] With the continuous growth of the number of cellular mobile communication users, especially in high-density urban areas, user distribution and service demand are highly uneven, and traditional average resource scheduling methods are difficult to balance overall throughput and edge user experience.

[0003] Based on this, related technologies have improved resource scheduling methods. These improved methods mainly include: edge user-priority-based resource scheduling methods, which identify edge users, such as those with low SINR (Signal to Interference plus Noise Ratio) or low RSRP (Reference Signal Receiving Power), and assign them high priority for resource scheduling; and adaptive resource scheduling methods based on cell load prediction, which predict the overall cell load trend and adjust resource allocation strategies accordingly. However, these resource scheduling methods still lack comprehensive perception and dynamic adaptive capabilities regarding local hotspot load distribution and edge user needs. Summary of the Invention

[0004] This invention was developed to at least partially address the technical problem that existing resource scheduling methods lack comprehensive perception and dynamic adaptive capabilities regarding the distribution of local hotspot loads and the needs of edge users.

[0005] According to one aspect of the present invention, an artificial intelligence-based resource scheduling method is provided, comprising: The cell coverage area is divided into multiple grids, and the grid weight, number of users, average service pressure of users in the grid, and actual grid load of each grid are obtained at the current time and at a preset number of historical time steps. The grid weight represents the degree to which the priority of users in the grid amplifies the grid load prediction, and the service pressure represents the degree of tension of user resource demand. Based on the grid weight, number of users, average service pressure of users within the grid, and actual grid load of each grid at the current time and a preset number of historical time steps, an LSTM-based grid load prediction model is used to output the grid load prediction value at the start of the next scheduling cycle. Based on the proportion of each grid's grid priority score in the total grid priority scores of all grids, the total bandwidth resources of the cell are allocated to each grid proportionally to obtain the schedulable resource amount of each grid in the current scheduling cycle; wherein the grid priority score is calculated based on the grid load prediction value and the user weight of each user in the grid, and represents the priority of the grid in resource allocation. For each grid, the schedulable resources of the grid in the current scheduling period are allocated to each user according to the proportion of each user's weight in the total user weight of all users in the grid; wherein the user weight represents the user's priority in resource scheduling.

[0006] Optionally, obtaining the grid weights includes: Based on the service pressure of each user in the grid, the distance between each user and the antenna of the cell they access, and the maximum coverage of the cell, the priority score of each user in the grid is calculated using a preset user priority scoring formula. The number of users within a grid is obtained based on a preset rasterization function. The grid weight is calculated based on the priority score of each user within the grid and the number of users within the grid, using a preset grid weight calculation formula.

[0007] Optionally, the user priority scoring formula is: ; in, To assign a priority score to the i-th user within the grid at time t; This is the preset first adjustment coefficient; This is the preset second adjustment coefficient; This refers to the service pressure metric for the i-th user within the grid at time t. Let be the distance between the i-th user and the antenna of the cell they are connected to within the grid at time t; This represents the maximum coverage area of ​​the residential community. The formula for calculating the grid weight is: ; in, Let (x, y) be the grid weight at time t, and (x, y) represent the coordinates of the grid's reference point. These are preset weighting coefficients; To assign a priority score to the i-th user within the grid at time t; Let be the set of users within the grid at time t; Let be the number of users in the grid at time t.

[0008] Optionally, obtaining the average service pressure of users within the grid includes: The service pressure index for each user is calculated based on the ratio of the data transmission rate required by each user within the grid to its achievable rate; wherein the achievable rate represents the maximum data transmission rate that the network can provide to the user. The average service pressure of users within a grid is calculated by dividing the sum of the service pressure indicators of all users within the grid by the number of users within the grid.

[0009] Optionally, the grid load prediction model is trained in the following manner: Multiple training samples are generated by a sliding window. Each training sample includes a time series consisting of grid weights, number of users, average service pressure of users in the grid, and actual load of the grid for a preset number of historical time steps. Obtain the label corresponding to each training sample, where the label represents the actual grid load at the next time step of the time window corresponding to the training sample. Based on the multiple training samples and the label corresponding to each training sample, the initial grid load prediction model based on LSTM is trained to obtain the trained grid load prediction model.

[0010] Optionally, the grid priority score is calculated using the following formula: ; in, For at any time The grid priority score is given, where (x, y) represents the coordinates of the grid's reference point, and t is the current time. Indicates the start time of the next scheduling cycle; For at any time The grid resource surplus index characterizes the degree of resource idleness assessed based on the grid load forecast value; Let be the user weight of the i-th user in the grid at time t; Let be the set of users within the grid at time t.

[0011] Optionally, the grid resource surplus index is calculated using the following formula: ; in, For at any time The grid resource surplus index; max(·) is the maximum value function; The preset target load threshold characterizes the ideal load on the mesh; This is the predicted grid load value at the start of the next scheduling cycle.

[0012] Optionally, the user weight is calculated using the following formula: ; in, Let be the user weight of the i-th user in the grid at time t; To assign a priority score to the i-th user within the grid at time t, The preset power exponent; This is the preset third adjustment coefficient; Let be the resource gap parameter for the i-th user in the grid at time t, representing the shortage of user resource demand; The data transmission rate required by the i-th user within the grid at time t; To prevent zero factor; This is the preset fourth adjustment coefficient; For at any time The grid resource surplus index.

[0013] Optionally, the resource gap parameter is calculated using the following formula: ; in, Let be the resource gap parameter for the i-th user in the grid at time t; max(·) is the maximum value function; The data transmission rate required by the i-th user within the grid at time t; Let be the achievable rate of the i-th user within the grid at time t, and let be the maximum data transmission rate that the network can provide to the i-th user at time t.

[0014] According to another aspect of the present invention, an artificial intelligence-based resource scheduling device is provided, comprising: The grid data acquisition module is configured to divide the cell coverage area into multiple grids and acquire the grid weight, number of users, average service pressure of users within the grid, and actual grid load of each grid at the current time and at a preset number of historical time steps. The grid weight represents the degree to which the priority of users within the grid amplifies the grid load prediction, and the service pressure represents the degree of tension in user resource demand. The grid load prediction module is configured to output the grid load prediction value at the start of the next scheduling cycle based on the grid weight, number of users, average service pressure of users in the grid and actual grid load of each grid at the current time and a preset number of historical time steps. The grid resource allocation module is configured to allocate the total bandwidth resources of the cell to each grid according to the proportion of the grid priority score of each grid in the total grid priority score of all grids, so as to obtain the amount of schedulable resources for each grid in the current scheduling cycle; wherein the grid priority score is calculated based on the grid load prediction value and the user weight of each user in the grid, and represents the priority of the grid in resource allocation. The user resource scheduling module is configured to allocate the schedulable resources of each grid to each user in the current scheduling cycle according to the proportion of each user's user weight in the total user weight of all users in the grid; wherein the user weight represents the user's priority in resource scheduling.

[0015] According to another aspect of the present invention, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the aforementioned resource scheduling method based on artificial intelligence.

[0016] According to another aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, wherein when the computer program is executed by a processor, the processor performs the aforementioned artificial intelligence-based resource scheduling method.

[0017] The technical solution provided by this invention may include the following beneficial effects: The artificial intelligence-based resource scheduling method and apparatus provided by this invention subdivides the cell coverage area into multiple grids and acquires the weight, number of users, average service pressure, and actual load of each grid in real time, thereby spatially perceiving the distribution of local hotspots and edge users. It utilizes an LSTM model to predict future grid loads based on current and historical data, enabling proactive resource allocation and anticipating potential local overloads. Total bandwidth resources are allocated proportionally to each grid through grid priority scoring, and then resources are allocated to specific users within each grid according to user weights, forming a two-level allocation framework of grid-level resource allocation and intra-grid user resource scheduling. This scheme can dynamically perceive changes in local load distribution and edge user demand, and adaptively adjust resource allocation, effectively overcoming the shortcomings of related technologies that cannot simultaneously handle local hotspots, edge users, and dynamic load changes.

[0018] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the description, claims, and drawings. Attached Figure Description

[0019] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the technical solutions of the present invention, and do not constitute a limitation on the technical solutions of the present invention.

[0020] Figure 1 A flowchart illustrating a resource scheduling method based on artificial intelligence provided in an embodiment of the present invention; Figure 2 A flowchart illustrating another resource scheduling method based on artificial intelligence provided in an embodiment of the present invention; Figure 3 A schematic diagram of the structure of an artificial intelligence-based resource scheduling device provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the specific implementation methods of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific implementation methods described herein are for illustration and explanation only and are not intended to limit the present invention.

[0022] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a set order or sequence; furthermore, in the absence of conflict, the embodiments and features in the embodiments of this invention can be arbitrarily combined with each other. In the following description, the use of suffixes such as "module," "component," or "unit" to represent elements is only for the convenience of the description of this invention and has no inherent meaning. Therefore, "module," "component," or "unit" can be used interchangeably.

[0023] The relevant technologies mainly employ resource scheduling methods based on edge user priority and adaptive resource scheduling methods based on cell load prediction, which have the following specific technical problems: Insufficient response to local hotspots and edge users: Most scheduling algorithms only consider the average load of the cell and ignore the uneven distribution of users, resulting in local congestion and a decline in the user experience at the edge.

[0024] Ignoring local load distribution: Existing scheduling algorithms mostly allocate resources at the cell level, which cannot detect and effectively solve the problem of local resource shortage.

[0025] Lack of dynamic adaptive capability: Some scheduling algorithms adopt a fixed priority strategy, such as providing resources to high-priority users or specific services. They cannot dynamically adjust according to real-time load changes, local overload conditions or fluctuations in user behavior, thus limiting resource utilization.

[0026] To address the aforementioned problems, this invention provides a resource scheduling scheme that combines user behavior awareness, edge user priority, gridded load prediction, and dynamic scheduling to simultaneously improve edge user experience and overall resource utilization. Specific embodiments are described in detail below.

[0027] Figure 1 This is a flowchart illustrating a resource scheduling method based on artificial intelligence, provided as an embodiment of the present invention. Figure 1 As shown, the method includes the following steps S101 to S104.

[0028] S101. Divide the cell coverage area into multiple grids, and obtain the grid weight, number of users, average service pressure of users in the grid, and actual load of the grid for each grid at the current time and at a preset number of historical time steps.

[0029] The grid weight represents the degree to which user priority within the grid amplifies the grid load prediction, and the service pressure represents the degree of tension in user resource demand.

[0030] In this step, the cell coverage area is divided into multiple fixed-size grids, which are generally square, with each grid representing a sub-region. Of course, due to the shape of the cell coverage area, the grids corresponding to the sub-regions at its perimeter may not be complete.

[0031] S102. Based on the grid weight, number of users, average service pressure of users in the grid and actual grid load of each grid at the current time and a preset number of historical time steps, a grid load prediction model based on LSTM is adopted to output the grid load prediction value at the start of the next scheduling cycle.

[0032] LSTM (Long Short-Term Memory) is a type of recurrent neural network (RNN) in deep learning and is one of the core technologies of artificial intelligence. It effectively solves the gradient vanishing and exploding problems of traditional RNNs by introducing gating mechanisms (forget gate, input gate, output gate), enabling it to learn long-term dependencies in time series. In load forecasting tasks, LSTM's advantages lie in its ability to automatically extract nonlinear time-series features from historical data, adapt to dynamically changing patterns, and support multivariate inputs (such as load, user density, service pressure, etc.), thereby achieving high-precision future state prediction.

[0033] S103. Based on the proportion of each grid's grid priority score in the total grid priority score of all grids, allocate the total bandwidth resources of the cell to each grid proportionally to obtain the amount of schedulable resources for each grid in the current scheduling cycle.

[0034] The grid priority score is calculated based on the grid load prediction value and the user weight of each user in the grid, and represents the priority of the grid in resource allocation.

[0035] S104. For each grid, based on the proportion of each user's weight in the total user weight of all users in the grid, allocate the schedulable resources of the grid to each user in the current scheduling period.

[0036] The user weight represents the user's priority in resource scheduling.

[0037] In this embodiment, the cell coverage area is subdivided into multiple grids, and the weight, number of users, average service pressure, and actual load of each grid are obtained in real time, thereby spatially perceiving the distribution of local hotspots and edge users. An LSTM model is used to predict the grid load at future moments based on current and historical data, enabling proactive resource allocation and anticipating potential local overloads. Total bandwidth resources are allocated proportionally to each grid through grid priority scoring (combining predicted load and user weight), and then resources are allocated to specific users within each grid according to user weight, forming a two-level allocation framework of grid-level resource allocation and intra-grid user resource scheduling. This scheme can dynamically perceive changes in local load distribution and edge user demand, and adaptively adjust resource allocation, effectively overcoming the shortcomings of related technologies that cannot simultaneously handle local hotspots, edge users, and dynamic load changes.

[0038] In one specific implementation, step S101 obtains the grid weights, which specifically includes the following steps S1011 to S1013.

[0039] S1011. Based on the service pressure of each user in the grid, the distance between each user and the antenna of the accessed cell, and the maximum coverage of the cell, the priority score of each user in the grid is calculated using a preset user priority scoring formula; S1012. Obtain the number of users within the grid based on the preset rasterization function; S1013. Based on the priority score of each user in the grid and the number of users in the grid, the grid weight is calculated using a preset grid weight calculation formula.

[0040] Further, in step S1012, the rasterization function is a mathematical function that maps continuous user position data to discrete grids and is used to count the number of users within the grids.

[0041] Specifically, (x, y) represents the reference point coordinates of the grid. For example, the lower left corner of the grid can be used as the reference point; (x i , y i ) represents the actual position coordinates of the i-th user. Assuming the grid is square with side length L, when x ≤ x i ≤ x + L and y ≤ y i ≤ y + L, define δ(x - x i , y - y i ) = 1, indicating that user i falls within the grid (x, y); otherwise, that is, x i < x, x i > x + L, y i < y and y i > y + L, satisfying any one of them, define δ(x - x i , y - y i ) = 0, indicating that user i does not fall within the grid (x, y). Then, count the number of users within the grid (x, y) , and .

[0042] In this embodiment, by first calculating the priority score of each user within the grid (combining service pressure, distance, etc.) and then aggregating to obtain the grid weight, the influence of edge users (high-score users) is amplified at the grid level, so that in subsequent load prediction and resource allocation, grids with high-priority user aggregations are given priority attention, improving the response ability to the needs of edge users.

[0043] In a specific implementation manner, in step S101a, the user priority score formula is: .

[0044] Where is the priority score of the i-th user within the grid at time t; is a preset first adjustment coefficient; is a preset second adjustment coefficient; is the service pressure index of the i-th user within the grid at time t; is the distance between the i-th user within the grid and the antenna of the accessed cell at time t; is the maximum coverage range of the cell. Specifically, the maximum coverage range of the cell can be pre-configured according to the designed coverage radius of the cell, or the maximum value of the estimated distances of all users accessing this cell can be obtained through periodic statistics and used as the maximum coverage range of the cell.

[0045] In this embodiment, the user priority scoring formula integrates service pressure and distance normalization factors, simultaneously reflecting the user's resource scarcity and geographical disadvantage. Users with higher scores have higher priority, indicating a greater need for resource allocation. Users with priority scores exceeding a preset threshold can be defined as marginal users requiring priority resource allocation during resource scheduling.

[0046] In one specific embodiment, in step S101c, the formula for calculating the grid weight is: .

[0047] in, Let (x, y) be the grid weight at time t, and (x, y) represent the coordinates of the grid's reference point. These are preset weighting coefficients; To assign a priority score to the i-th user within the grid at time t, Let be the set of users within the grid at time t. This represents summing the priority scores of all users within the grid (x,y) at time t; Let be the number of users in the grid at time t.

[0048] In this embodiment, the weight of grids with high-priority rated users is amplified in the grid weight calculation formula, so that the load prediction model pays more attention to grids where edge users are clustered, thereby improving the prediction accuracy of local load.

[0049] In one specific implementation, step S101 obtains the average service pressure of users within the grid, specifically including the following steps S1014 and S1015.

[0050] S1014. Calculate the service pressure index for each user based on the ratio of the data transmission rate required by each user within the grid to its achievable rate.

[0051] The formula for calculating the service stress index is as follows: .

[0052] in, This refers to the service pressure metric for the i-th user within the grid at time t. This represents the data transmission rate required by the i-th user within the grid at time t, in Mbps or kbps, reflecting the intensity of the user's demand for network resources, such as the need for higher speeds when watching high-definition video, playing games, or downloading. Let be the achievable rate of the i-th user in the grid at time t, expressed in Mbps or kbps. It represents the maximum data transmission rate that the network can provide to the user, that is, the theoretical upper limit of the data transmission rate that the user can achieve under the current channel quality and resource allocation.

[0053] S1015. Calculate the average service pressure of users within the grid based on the ratio of the sum of service pressure indicators of all users within the grid to the number of users within the grid.

[0054] The formula for calculating average service pressure is as follows: .

[0055] in, Let be the average service pressure of users within grid (x,y) at time t; Let be the number of users in grid (x,y) at time t; Let be the set of users within the grid (x,y) at time t; Let be the service pressure metric for the i-th user within the grid at time t, and This represents the summation of the service pressure metrics for all users within the grid (x,y) at time t.

[0056] In this embodiment, the service pressure indicators at the user level are aggregated into the average service pressure at the grid level. This can eliminate the impact of random fluctuations in individual users, accurately reflect the overall resource stress of the grid, and provide stable and representative input features for the LSTM load prediction model, thereby improving the reliability of the prediction.

[0057] In one specific implementation, step S101 obtains the actual load of the grid, specifically including the following step S1015.

[0058] S1015. Calculate the actual load on the grid based on the data transmission rate required by each user within the grid and the maximum allocatable capacity of the grid.

[0059] The formula for calculating the actual load on the grid is as follows: .

[0060] in, This represents the actual load on the grid (x,y) at time t; To determine the data transfer rate required by the i-th user within the grid at time t, Let be the set of users within the grid (x,y) at time t. This represents the summation of the data transfer rates required by all users within the grid (x,y) at time t; The maximum allocatable capacity of a grid (x,y) refers to the maximum total data transmission rate that all users in the grid can obtain under ideal conditions. It can be pre-allocated according to the ratio of the total cell bandwidth and the grid area.

[0061] In this embodiment, the normalized actual grid load (dimensionless) is obtained by calculating the ratio of the sum of the demand rates of all users within the grid to the maximum allocable capacity of the grid, which intuitively reflects the current resource occupancy rate of the grid. This load metric can be used as a historical label for training the LSTM model, enabling the prediction model to learn the true load evolution pattern. Compared with related technologies that directly use indirect indicators such as the number of users or average throughput to obtain the load, the load calculation method of this invention more accurately quantifies the resource stress of the grid, providing a reliable data foundation for subsequent load prediction, thereby improving the accuracy of overall scheduling decisions.

[0062] In one specific implementation, in step S102, the training method of the grid load prediction model includes the following steps S1021 to S1023.

[0063] S1021. Generate multiple training samples through a sliding window. Each training sample includes a time series consisting of grid weights, number of users, average service pressure of users within the grid, and actual load of the grid for a preset number of historical time steps. S1022. Obtain the label corresponding to each training sample, wherein the label is the actual grid load at the next moment of the time window corresponding to the training sample; S1023. Based on the multiple training samples and the label corresponding to each training sample, train the initial grid load prediction model based on LSTM to obtain the trained grid load prediction model.

[0064] In this process, multiple training samples are generated using a sliding window. These samples have overlapping features. For example, three training samples are generated, with a window length (i.e., the number of consecutive time steps in the input sequence) of 3. (each) This includes a time series consisting of grid weights, number of users, average service pressure per user within the grid, and actual grid load at a given time point, labeled L. t (i.e., the actual grid load at time t). The corresponding label is L t-1 ; The corresponding label is L t-2。

[0065] After training the grid load prediction model, the model outputs the predicted grid load value at the start of the next scheduling cycle, specifically: .

[0066] In this embodiment, a sliding window approach is used to generate training samples. Each sample contains feature sequences from multiple consecutive historical time steps, labeled with the actual load at the next moment. This training method enables the LSTM model to learn dynamic patterns (such as trends and periodicity) of load changes over time, thereby accurately predicting future loads based on currently known data in practical applications and achieving adaptive look-ahead scheduling.

[0067] In one specific implementation, in step S103, the grid priority score is calculated using the following formula: .

[0068] in, For at any time The grid priority score is given, where (x, y) represents the coordinates of the grid's reference point, and t is the current time. It indicates the start time of the next scheduling cycle, which is equivalent to the end time of the current scheduling cycle; For at any time The grid resource surplus index characterizes the degree of resource idleness assessed based on the grid load forecast value; Let be the user weight of the i-th user in the grid at time t. Let be the set of users within the grid at time t. This represents summing the user weights of all users within the grid (x,y) at time t.

[0069] Accordingly, step S103 calculates the schedulable resource quantity for each grid in the current scheduling period using the following formula: .

[0070] in, This represents the amount of schedulable resources in the current scheduling period grid (x, y); For at any time The grid priority score for the grid (x,y); For at any time The grid priority score is calculated for a grid (m,n), where the cell coverage area is divided into M×N grids. Indicates the time Sum the grid priority scores of all grids within the community; This represents the total bandwidth resources available for allocation within the cell during the current scheduling period.

[0071] In this embodiment, the calculation formula for grid priority scoring multiplies the grid resource surplus index (based on load prediction value) by the weight sum of all users in the grid. The surplus index reflects the resource idleness, and the user weight sum reflects the overall urgency of grid demand. The priority index determined by the combination of the two can dynamically guide the allocation of resources among grids: idle grids or high-demand grids are given higher priority, thereby improving resource utilization efficiency and local overload mitigation capabilities.

[0072] In one specific implementation, the grid resource surplus index is calculated using the following formula: .

[0073] in, For at any time The grid resource surplus index; max(·) is the maximum value function; The preset target load threshold characterizes the ideal load on the mesh; This is the predicted grid load value at the start of the next scheduling cycle.

[0074] In this embodiment, the grid resource surplus index uses the ratio of predicted load to target load threshold, outputting a non-negative value through a maximum value function. The index is positive (resources are surplus) when the predicted load is lower than the target load threshold; otherwise, it is zero (resources are scarce). This index quantifies the future availability of resources in a concise mathematical form, providing a clear priority basis for grid-level resource allocation and avoiding localized overload caused by average allocation.

[0075] In one specific implementation, in step S104, the user weight is calculated using the following formula: .

[0076] in, Let be the user weight of the i-th user in the grid at time t; To assign a priority score to the i-th user within the grid at time t, The preset power exponent; This is the preset third adjustment coefficient; Let be the resource gap parameter for the i-th user in the grid at time t, representing the shortage of user resource demand; The data transmission rate required by the i-th user within the grid at time t; To prevent zero factor; This is the preset fourth adjustment coefficient; For at any time The grid resource surplus index.

[0077] Accordingly, step S104 uses the following formula to calculate the amount of resources allocated to each user in the grid during the current scheduling period: .

[0078] in, The amount of resources allocated to the i-th user within the grid during the current scheduling period; Let be the user weight of the i-th user in the grid at time t; Let be the user weight of the j-th user in the grid at time t. Let be the set of users within the grid (x,y) at time t. This represents summing the user weights of all users within the grid (x, y) at time t; This represents the amount of schedulable resources in the current scheduling period grid (x, y).

[0079] In this embodiment, the user weight formula integrates three factors: edge user score (reflecting user priority), resource gap factor (reflecting the degree of resource shortage for users), and grid resource surplus factor (reflecting the degree of resource idleness in the grid). The weight of each user is dynamically adjusted through the product of the three factors, so that users with high demand, high risk, and located in idle grids receive a higher proportion when allocating resources within the grid, thus achieving refined and fair scheduling at the user level.

[0080] In one specific implementation, the resource gap parameter is calculated using the following formula: .

[0081] in, Let be the resource gap parameter for the i-th user in the grid at time t; max(·) is the maximum value function; The data transmission rate required by the i-th user within the grid at time t; Let be the achievable rate of the i-th user within the grid at time t, and let be the maximum data transmission rate that the network can provide to the i-th user at time t.

[0082] In this embodiment, the resource gap parameter directly calculates the difference between the user's demand rate and the achievable rate, quantifying the absolute amount of user resource shortage. It can intuitively reflect the user's current urgency and provide an accurate and calculable input for the gap factor in the user weight formula, ensuring that the scheduling decision responds to the real demand.

[0083] The AI-based resource scheduling method provided in this invention achieves comprehensive perception and dynamic adaptation capabilities for local hotspot load distribution and edge user demand by gridding the cell coverage area and introducing LSTM-based load prediction and a collaborative scheduling mechanism composed of grid priority scores and user weights. Specifically, user distribution, service pressure, and actual load are statistically analyzed using grids as basic units, enabling the system to accurately locate local hotspots and edge user clusters, avoiding the average blind spots of traditional cell-level scheduling. The LSTM model learns the mapping relationship between historical time series (grid weights, number of users, average service pressure, and actual load) and future load, outputting the grid load prediction value at the start of the next scheduling cycle. This allows resource allocation to anticipate upcoming peaks, transforming scheduling from a passive response to proactive defense. Based on this, the first-level scheduling allocates total bandwidth resources to each grid proportionally according to the grid priority score (determined by the resource surplus index derived from the predicted load and the sum of user weights within the grid), achieving differentiated quotas for different grids. The second-level scheduling allocates resources within each grid to specific users according to their weights, thus spatially tilting towards local hotspots and prioritizing high-demand or edge users at the user level. Meanwhile, all key indicators (service pressure, priority score, grid weight, user weight, resource gap, grid priority score, etc.) are dynamically calculated based on real-time or predicted data, and adjustment coefficients can be preset to adapt to different network environments. Therefore, the scheduling strategy can be automatically adjusted according to changes in user distribution, service demand, and channel quality. This systematically solves the problems of insufficient perception of local hotspot load distribution and edge user demand, and lack of dynamic adaptive capability in resource scheduling methods in related technologies. While effectively improving the edge user experience, it also optimizes the overall resource utilization efficiency.

[0084] Figure 2 This is a flowchart illustrating another resource scheduling method based on artificial intelligence provided in an embodiment of the present invention. Figure 2 As shown, the method includes the following steps S201 to S204.

[0085] S201. Acquisition of user behavior perception data.

[0086] The purpose of this step is to obtain real-time user distribution and service pressure information within the cell, providing accurate and dynamic input data for subsequent edge user identification, load prediction, and resource optimization. Specifically, by monitoring the current user status in the cell and calculating user service pressure, a map of the current user distribution and service pressure within the cell is generated to determine which areas are densely populated and which users may face service pressure.

[0087] The inputs for this step include each user's real-time location, each user's current data transmission rate requirements, and each user's current available resources.

[0088] First, the location of users in continuous space is discretized by a rasterization function to generate a quantifiable user density matrix, which can be used for statistical analysis, visualization, and subsequent AI (Artificial Intelligence) prediction.

[0089] Specifically, rasterization refers to dividing the cell coverage area into several grids (similar to a checkerboard pattern), with each grid representing a sub-region. This is used to statistically analyze user distribution and load within that sub-region. The goal is to map continuous, dispersed user location data onto discrete grids to form a user density matrix for the corresponding cell, facilitating quantitative statistics and modeling. This matrix will then be used for cell load analysis and prediction.

[0090] The specific operation of rasterization is to divide the cell coverage area into multiple fixed-size grids, for example, each grid is 10m × 10m. Based on the actual coordinates (x, y, y) of user i... i ,y i This determines which grid a user falls into and counts the number of users in each grid.

[0091] A rasterization function is a mathematical function that maps continuous user location data to a discrete grid, used to count the number of users within each grid.

[0092] The specific formula for the rasterization function is as follows: .

[0093] Among them, (x i ,y i (x, y) represents the actual position coordinates of the i-th user; (x, y) represents the coordinates of the reference point of the grid, which can be the lower left corner of the grid; δ ) is a rasterization function that indicates whether user i falls within the grid (x,y): assuming the grid is a square with side length L, when x ≤ x i ≤x+L and y≤y i When ≤y+L, define δ(xx) i yy i If δ(xx) = 1, it means that user i falls within the grid (x, y). Otherwise, define δ(xx) = 1. i yy i If )=0, it means that user i does not fall within the grid (x,y); This represents the number of users who fall within the grid (x, y) at time t, where t is the current time.

[0094] Calculate the service pressure index at time t for each user within the grid. : .

[0095] in, This represents the data transmission rate (i.e., demand rate) required by the i-th user at time t, in Mbps or kbps, reflecting the intensity of the user's demand for network resources, such as the need for higher speeds when watching high-definition videos, playing games, or downloading. This represents the achievable rate of the i-th user at time t, in Mbps or kbps. The achievable rate characterizes the maximum data transmission rate that the network can provide to the user, that is, the upper limit of the data transmission rate that the user can theoretically achieve under the current channel quality and resource allocation.

[0096] Specifically, the achievable rate can be calculated from the number of PRBs allocated to the user by the base station and the user's real-time channel quality: .

[0097] in, This represents the number of Physical Resource Blocks (PRBs) allocated to the i-th user by the network at time t, reflecting the amount of resources obtained by that user; This represents the real-time spectral efficiency (in bps / Hz) of the i-th user at time t, which can be obtained by mapping the channel quality indicator (CQI) reported by the user or the directly measured signal-to-interference-plus-noise ratio (SINR). It reflects the quality of the wireless channel in real time; the better the channel quality, the higher the spectral efficiency. This represents the bandwidth (in Hz) of each Physical Resource Block (PRB), which is a fixed system parameter. For example, in LTE (Long Term Evolution), the bandwidth of a PRB is 180 kHz; in 5G NR, it can be configured according to the subcarrier spacing.

[0098] Used to measure whether available resources can meet user needs. A value greater than 1 indicates that user demand exceeds available resources, resulting in service pressure. ≤1 indicates that available resources can meet user needs.

[0099] Next, obtain the actual load of each grid at time t. This serves as the load label for subsequent load forecasting.

[0100] .

[0101] in, This represents the actual load on the grid (x,y) at time t; To determine the data transfer rate required by the i-th user within the grid at time t, Let be the set of users who fall within the grid (x,y) at time t. This represents the summation of the data transfer rates required by all users within the grid (x,y) at time t; The maximum allocatable capacity of a grid (x,y) refers to the maximum total data transmission rate that all users in the grid can obtain under ideal conditions (i.e., the upper limit of the grid's capacity), which can be pre-allocated according to the ratio of the total cell bandwidth to the grid area.

[0102] S202. Edge User Identification and Priority Scoring.

[0103] This step identifies "marginal users" within the cell who may have low service quality and provides a priority reference for their resource scheduling. Specifically, it identifies users who are far from the cell and under high resource pressure, so that their service quality can be prioritized in subsequent scheduling.

[0104] Define the user priority scoring formula: .

[0105] in, α、β : Adjustment factor, used to balance the weight of service pressure and geographical location; : The distance between the i-th user and the antenna of the cell they are connected to at time t; The maximum coverage area of ​​a cell is used to normalize distances to 0-1. It can be pre-configured based on the cell's designed coverage radius, specifically during system deployment. Alternatively, it can be calculated by periodically collecting the estimated distances of all users accessing the cell and taking the maximum value as the maximum coverage area. The estimated distance of users can be estimated using time advance (TA), i.e., D. i,估计 ≈(TA×78) meters; : The priority score of the i-th user at time t, used for sorting and priority resource allocation. The larger the value, the more priority the user needs to guarantee resources.

[0106] In this step, The larger the value, the higher the pressure on user services; The larger the value, the closer the user is to the edge of the community.

[0107] Then, filter. Higher-level users, such as Users whose scores exceed the preset threshold are considered edge users who require priority resource allocation during resource scheduling.

[0108] S203. Predict cell load.

[0109] The core objective of this step is to predict load changes in each sub-area of ​​the cell in the near future, enabling proactive resource scheduling. Related technologies typically rely on historical averages or closed-loop feedback for resource scheduling, which cannot quickly respond to sudden user surges or changes in traffic volume. This step uses AI prediction to allow the network to adjust resources in advance, avoiding overload or degraded user experience at the edge. It also focuses on considering the impact of high-risk edge users on local load, avoiding the oversight of load peaks caused by hotspot areas or edge user clusters in averaged predictions.

[0110] For each grid, combine the number of users, average service pressure, and actual load over the current and past k time steps into a time series: .

[0111] in, Let be the number of users who fall within the grid (x, y) at time t; Let be the average service pressure of users within grid (x,y) at time t, and ; Let be the actual load on the grid (x,y) at time t.

[0112] Each time step includes not only the number of users, but also the service stress metric P. i The average value and the actual load provide local load pressure information for subsequent load prediction models.

[0113] User priority rating S i Convert to grid weight W x,y To enhance the impact of high-risk users on load forecasting: .

[0114] in, The set of users who fall within the grid (x,y) at time t; User priority rating at time t; The number of users falling within the grid (x,y) at time t; γ: Weighting coefficient, which controls the degree of influence of edge users on the prediction.

[0115] This indicates that at time t, the grid (x, y) is amplified and focused on due to the presence of high-rated edge users.

[0116] By amplifying the impact of edge users on local load forecasting through weighting, the load forecasting not only reflects the average user distribution but also the potential high-risk local load peaks.

[0117] Predicting grid load using an LSTM model: .

[0118] The curly braces represent sequences. The time series at each time step is first multiplied by grid weights, and then the weighted sequence is input into the LSTM prediction model.

[0119] The load prediction model amplifies edge user data through grid weights, making local load peaks more significant.

[0120] The prediction results from each grid are combined into a future cell load prediction matrix: .

[0121] The coverage area of ​​the cell is divided into M×N grids.

[0122] In this step, the user priority score S generated in step S202 i The information used to form the grid weights W x,y This enhances the load forecasting model's focus on high-risk users. The load forecasting results reflect the predicted load of each grid, providing not only the overall average load but also capturing load spikes that may be caused by local edge users, thus ensuring that subsequent resource scheduling steps in S204 are more targeted.

[0123] S204. Adaptive Resource Optimization and Scheduling.

[0124] The purpose of this step is to prioritize high-risk users (high-priority score users) within a given total cell resource budget, proactively prevent predicted local peaks, and ensure overall throughput and fairness. Resource allocation primarily involves PRBs (or bandwidth slots) and transmit power.

[0125] To allocate resources more precisely, resources need to be allocated first at the spatial level (all grids), and then specific resources need to be allocated within each grid according to user weights. Therefore, the overall framework is a two-level allocation.

[0126] The first level is to allocate grid-level slicing quotas (spatial quotas): allocating a schedulable resource quota to each grid based on "predicted load × edge weight".

[0127] The second level is user-level scheduling (PRB / power distribution): within each grid, PRBs and power are distributed to specific users according to a weighted proportional fairness principle.

[0128] (1) Grid-level slice quota: Definition B total This refers to the total bandwidth resources available for allocation in a given cell within a specific scheduling period, typically measured by the number of physical resource blocks or spectral bandwidth. In the previous step, the prediction time interval Δt is set to the same as the resource scheduling period, for example, Δt = 1 millisecond. The future load predicted by the LSTM model corresponds to the resource demand in the current scheduling period, such as the load predicted by the LSTM model at time t+Δt, which guides resource allocation in the current scheduling period (from t to t+Δt).

[0129] For user i within the grid (x,y), a resource gap parameter is defined, which refers to the difference between the resources actually needed by the user (demand rate) and the available resources (reachability rate) that the current system can provide to the user:

[0130] There is no resource gap (gap=0) if the user demand rate is less than or equal to the reachable rate; a resource gap only exists when the demand rate is greater than the reachable rate.

[0131] Define grid resource surplus indicators: .

[0132] In this step, the grid resource surplus index is used to measure the degree to which resource demand in a grid is lower than available resources (i.e., resource surplus). It reveals whether network resources are redundant at a certain time.

[0133] Define L target The target load threshold represents the ideal load level that each grid is expected to maintain. The cell network coverage area is divided into multiple grids, each containing a certain number of users and resources. Predicted load is calculated separately for each grid to facilitate local scheduling and optimization, aiming to avoid local overload while ensuring reasonable resource utilization.

[0134] In this formula, if the predicted load is less than the target load threshold, the ratio > 1 → the system has resource redundancy; if the predicted load is greater than the target load threshold, the ratio < 1 → the system has resource shortage and no redundancy.

[0135] Based on this formula, a non-negative resource surplus index is obtained: greater than 0 → resource redundancy, equal to 0 → resource shortage or overload.

[0136] Define user weight w iThis is introduced to measure the importance or urgency of each user when allocating resources within the grid. It integrates the user's own needs, resource gaps, and the load of the grid. The calculation formula can be expressed as: .

[0137] Among them, S i User priority rating; ρ: Power exponent, controlling S i Degree of influence; μ: adjustment coefficient, controlling the intensity of the gap's influence on the weight; ε: a small constant to prevent division by zero; : Adjustment coefficient, which controls the impact of grid load on user weights.

[0138] In short, w i =Edge user importance × User resource gap factor × Grid resource idle factor.

[0139] For each grid cell (x, y), calculate the grid priority score: .

[0140] Next, based on the grid priority score, the total resources B of the cell will be allocated. total The schedulable resources for each grid are allocated proportionally based on the grid priority score, and the amount of schedulable resources obtained by each grid in the current scheduling cycle is: .

[0141] (2) User-level scheduling.

[0142] Within each grid, resources allocated to that grid are finely distributed to specific users, taking into account individual user differences (demand gaps, priorities, etc.) to ensure that high-demand / high-risk users receive resources first, while also ensuring fairness. Specifically, the Weighted Proportional Fair (W-PF) principle is used to allocate resources for the current scheduling cycle: .

[0143] w i Larger users (e.g., those with large resource gaps, high grid loads, or VIP users) are given priority for more resources. Weighting ensures local fairness: each user receives resources proportional to their needs and risks, rather than simply on average. The PRB or power allocated to each user can be directly used for scheduling and distribution, enabling dynamic resource optimization.

[0144] In summary, the first-level grid resource allocation allocates total resources based on spatial hotspots (grids) to address local overload issues; the second-level user resource scheduling allocates resources within a grid based on user weights to address uneven resource distribution among individual users. By combining these two approaches, the system can simultaneously prioritize both spatial and user needs.

[0145] The artificial intelligence-based resource scheduling method provided in this invention first divides the network coverage area into multiple grids, performs future load prediction and calculates resource gap indicators for each grid, and identifies edge users and assigns them priority scores. Then, the total bandwidth resources of the cell are allocated to each grid according to grid priority, and then resource allocation is performed within each grid according to user weight (combining user priority scores, resource gaps, and grid resource surplus). This realizes adaptive resource scheduling of cellular networks based on gridded load prediction and user behavior priority. Resource scheduling prioritizes edge users, and the overall resource allocation is fair and efficient, thereby improving the edge user experience and alleviating local overload problems.

[0146] Figure 3 This is a schematic diagram of the structure of an artificial intelligence-based resource scheduling device provided in an embodiment of the present invention. Figure 3 As shown, the device includes: a grid data acquisition module 301, a grid load prediction module 302, a grid resource allocation module 303, and a user resource scheduling module 304.

[0147] The grid data acquisition module 301 is configured to divide the cell coverage area into multiple grids and acquire the grid weight, number of users, average service pressure of users within the grid, and actual grid load of each grid at the current time and at a preset number of historical time steps. The grid weight represents the degree to which the priority of users within the grid amplifies the grid load prediction, and the service pressure represents the degree of tension in user resource demand. The grid load prediction module 302 is configured to output the grid load prediction value at the start of the next scheduling cycle based on the grid weight, number of users, average service pressure of users in the grid and actual grid load of each grid at the current time and a preset number of historical time steps. The grid resource allocation module 303 is configured to allocate the total bandwidth resources of the cell to each grid according to the proportion of the grid priority score of each grid in the total grid priority score of all grids, so as to obtain the amount of schedulable resources for each grid in the current scheduling cycle; wherein the grid priority score is calculated based on the grid load prediction value and the user weight of each user in the grid, and represents the priority of the grid in resource allocation. The user resource scheduling module 304 is configured to allocate the schedulable resources of each grid to each user in the current scheduling cycle according to the proportion of the user weight of each user in the grid to the total user weight of all users in the grid; wherein the user weight represents the priority of the user in resource scheduling.

[0148] In one specific implementation, the grid data acquisition module 301 acquires the grid weight, specifically configured as follows: based on the service pressure of each user in the grid, the distance between each user and the antenna of the accessed cell, and the maximum coverage area of ​​the cell, the priority score of each user in the grid is calculated using a preset user priority scoring formula; based on a preset rasterization function, the number of users in the grid is acquired; based on the priority score of each user in the grid and the number of users in the grid, the grid weight is calculated using a preset grid weight calculation formula.

[0149] In one specific implementation, the user priority scoring formula is: ; in, To assign a priority score to the i-th user within the grid at time t; This is the preset first adjustment coefficient; This is the preset second adjustment coefficient; This refers to the service pressure metric for the i-th user within the grid at time t. Let be the distance between the i-th user and the antenna of the cell they are connected to within the grid at time t; This represents the maximum coverage area of ​​the residential community. The formula for calculating the grid weight is: ; in, Let (x, y) be the grid weight at time t, and (x, y) represent the coordinates of the grid's reference point. These are preset weighting coefficients; To assign a priority score to the i-th user within the grid at time t; Let be the set of users within the grid at time t; Let be the number of users in the grid at time t.

[0150] In one specific implementation, the grid data acquisition module 301 acquires the average service pressure of users within the grid, specifically configured as follows: calculating the service pressure index of each user based on the ratio of the data transmission rate required by each user within the grid to its achievable rate; wherein the achievable rate represents the maximum data transmission rate that the network can provide to the user; and calculating the average service pressure of users within the grid based on the ratio of the sum of the service pressure indices of all users within the grid to the number of users within the grid.

[0151] In one specific embodiment, the system further includes a model training module.

[0152] The model training module is specifically configured as follows: multiple training samples are generated through a sliding window. Each training sample includes a time series consisting of grid weights, number of users, average service pressure of users within the grid, and actual grid load for a preset number of historical time steps; the label corresponding to each training sample is obtained, and the label is the actual grid load at the next moment of the time window corresponding to the training sample; based on the multiple training samples and the label corresponding to each training sample, the initial grid load prediction model based on LSTM is trained to obtain the trained grid load prediction model.

[0153] In one specific implementation, the grid priority score is calculated using the following formula: ; in, For at any time The grid priority score is given, where (x, y) represents the coordinates of the grid's reference point, and t is the current time. Indicates the start time of the next scheduling cycle; For at any time The grid resource surplus index characterizes the degree of resource idleness assessed based on the grid load forecast value; Let be the user weight of the i-th user in the grid at time t; Let be the set of users within the grid at time t.

[0154] In one specific implementation, the grid resource surplus index is calculated using the following formula: ; in, For at any time The grid resource surplus index; max(·) is the maximum value function; The preset target load threshold characterizes the ideal load on the mesh; This is the predicted grid load value at the start of the next scheduling cycle.

[0155] In one specific implementation, the user weight is calculated using the following formula: ; in, Let be the user weight of the i-th user in the grid at time t; To assign a priority score to the i-th user within the grid at time t, The preset power exponent; This is the preset third adjustment coefficient; Let be the resource gap parameter for the i-th user in the grid at time t, representing the shortage of user resource demand; The data transmission rate required by the i-th user within the grid at time t; To prevent zero factor; This is the preset fourth adjustment coefficient; For at any time The grid resource surplus index.

[0156] In one specific implementation, the resource gap parameter is calculated using the following formula: ; in, Let be the resource gap parameter for the i-th user in the grid at time t; max(·) is the maximum value function; The data transmission rate required by the i-th user within the grid at time t; Let be the achievable rate of the i-th user within the grid at time t, and let be the maximum data transmission rate that the network can provide to the i-th user at time t.

[0157] The AI-based resource scheduling device provided in this invention achieves comprehensive perception and dynamic adaptation capabilities for local hotspot load distribution and edge user demand by gridding the cell coverage area and introducing LSTM-based load prediction and a collaborative scheduling mechanism composed of grid priority scores and user weights. Specifically, it uses grids as basic units to statistically analyze user distribution, service pressure, and actual load, enabling the system to accurately locate local hotspots and edge user clusters, avoiding the average blind spots of traditional cell-level scheduling. It utilizes an LSTM model to learn the mapping relationship between historical time series (grid weights, number of users, average service pressure, and actual load) and future load, outputting the grid load prediction value at the start of the next scheduling cycle. This allows resource allocation to anticipate upcoming peaks, transforming scheduling from a passive response to an active defense. Based on this, the first-level scheduling allocates total bandwidth resources to each grid proportionally according to the grid priority score (determined by the resource surplus index derived from the predicted load and the sum of user weights within the grid), achieving differentiated quotas for different grids. The second-level scheduling allocates resources within each grid to specific users according to their weights, thus spatially tilting towards local hotspots and prioritizing high-demand or edge users at the user level. Meanwhile, all key indicators (service pressure, priority score, grid weight, user weight, resource gap, grid priority score, etc.) are dynamically calculated based on real-time or predicted data, and adjustment coefficients can be preset to adapt to different network environments. Therefore, the scheduling strategy can be automatically adjusted according to changes in user distribution, service demand, and channel quality. This systematically solves the problems of insufficient perception of local hotspot load distribution and edge user demand, and lack of dynamic adaptive capability in resource scheduling methods in related technologies. While effectively improving the edge user experience, it also optimizes the overall resource utilization efficiency.

[0158] Based on the same technical concept, embodiments of the present invention also provide a computer device, such as... Figure 4 As shown, the computer device includes a memory 401 and a processor 402. The memory 401 stores a computer program. When the processor 402 runs the computer program stored in the memory 401, the processor 402 executes the aforementioned resource scheduling method based on artificial intelligence.

[0159] Based on the same technical concept, this embodiment of the invention also provides a computer-readable storage medium storing a computer program thereon, wherein when the computer program is executed by a processor, the processor executes the aforementioned resource scheduling method based on artificial intelligence.

[0160] In summary, the AI-based resource scheduling method, apparatus, computer equipment, and storage medium provided in this invention can perceive local hotspots and edge user demands in real time. Through forward-looking load prediction and two-level collaborative scheduling, it effectively alleviates local overload, improves the edge user experience, and significantly enhances overall resource utilization and dynamic adaptability. This method refines the scheduling granularity from cell to grid and uses LSTM load prediction to drive resource allocation and scheduling, thereby accurately matching resource supply and demand in both spatial and user dimensions. This overcomes the shortcomings of related technologies, such as delayed resource adjustment response and neglect of local load distribution.

[0161] It will be understood by those skilled in the art that all or some of the steps, systems, or apparatuses disclosed above, and their functional modules / units, can be implemented as software, firmware, hardware, or suitable combinations thereof. In hardware implementations, the division between functional modules / units mentioned in the above description does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit (ASIC). Such software may be distributed on a computer-readable medium, which may include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, it is well known to those skilled in the art that communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0162] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An artificial intelligence-based resource scheduling method, characterized by, include: The cell coverage area is divided into multiple grids, and the grid weight, number of users, average service pressure of users in the grid, and actual grid load of each grid are obtained at the current time and at a preset number of historical time steps. The grid weight represents the degree to which the priority of users in the grid amplifies the grid load prediction, and the service pressure represents the degree of tension of user resource demand. Based on the grid weight, number of users, average service pressure of users within the grid, and actual grid load of each grid at the current time and a preset number of historical time steps, an LSTM-based grid load prediction model is used to output the grid load prediction value at the start of the next scheduling cycle. Based on the proportion of each grid's grid priority score in the total grid priority scores of all grids, the total bandwidth resources of the cell are allocated to each grid proportionally to obtain the schedulable resource amount of each grid in the current scheduling cycle; wherein the grid priority score is calculated based on the grid load prediction value and the user weight of each user in the grid, and represents the priority of the grid in resource allocation. For each grid, the schedulable resources of the grid in the current scheduling period are allocated to each user according to the proportion of each user's weight in the total user weight of all users in the grid; wherein the user weight represents the user's priority in resource scheduling. 2.The artificial intelligence-based resource scheduling method of claim 1, wherein, Obtaining the grid weights includes: Based on the service pressure of each user in the grid, the distance between each user and the antenna of the cell they access, and the maximum coverage of the cell, the priority score of each user in the grid is calculated using a preset user priority scoring formula. The number of users within a grid is obtained based on a preset rasterization function. The grid weight is calculated based on the priority score of each user within the grid and the number of users within the grid, using a preset grid weight calculation formula. 3.The AI-based resource scheduling method of claim 2, wherein, The user priority scoring formula is as follows: ； in, To assign a priority score to the i-th user within the grid at time t; This is the preset first adjustment coefficient; This is the preset second adjustment coefficient; This refers to the service pressure metric for the i-th user within the grid at time t. Let be the distance between the i-th user and the antenna of the cell they are connected to within the grid at time t; This represents the maximum coverage area of ​​the residential community. The formula for calculating the grid weight is: ； in, Let (x, y) be the grid weight at time t, and (x, y) represent the coordinates of the grid's reference point. These are preset weighting coefficients; To assign a priority score to the i-th user within the grid at time t; Let be the set of users within the grid at time t; Let be the number of users in the grid at time t. 4.The AI-based resource scheduling method of claim 1, wherein, Obtaining the average service pressure of users within the grid includes: The service pressure index for each user is calculated based on the ratio of the data transmission rate required by each user within the grid to its achievable rate; wherein the achievable rate represents the maximum data transmission rate that the network can provide to the user. The average service pressure of users within a grid is calculated by dividing the sum of the service pressure indicators of all users within the grid by the number of users within the grid. 5.The AI-based resource scheduling method of claim 1, wherein, The grid load prediction model was trained in the following manner: Multiple training samples are generated by a sliding window. Each training sample includes a time series consisting of grid weights, number of users, average service pressure of users in the grid, and actual load of the grid for a preset number of historical time steps. Obtain the label corresponding to each training sample, where the label represents the actual grid load at the next time step of the time window corresponding to the training sample. Based on the multiple training samples and the label corresponding to each training sample, the initial grid load prediction model based on LSTM is trained to obtain the trained grid load prediction model. 6.The artificial intelligence-based resource scheduling method of claim 1, wherein, The grid priority score is calculated using the following formula: ； in, For at any time The grid priority score is given, where (x, y) represents the coordinates of the grid's reference point, and t is the current time. Indicates the start time of the next scheduling cycle; For at any time The grid resource surplus index characterizes the degree of resource idleness assessed based on the grid load forecast value; Let be the user weight of the i-th user in the grid at time t; Let be the set of users within the grid at time t.

7. The artificial intelligence-based resource scheduling method of claim 6, wherein, The grid resource surplus index is calculated using the following formula: ； in, For at any time The grid resource surplus index; max(·) is the maximum value function; The preset target load threshold characterizes the ideal load on the mesh; This is the predicted grid load value at the start of the next scheduling cycle. 8.The AI-based resource scheduling method of claim 1 or 6, wherein, The user weight is calculated using the following formula: ； in, Let be the user weight of the i-th user in the grid at time t; To assign a priority score to the i-th user within the grid at time t, The preset power exponent; This is the preset third adjustment coefficient; Let be the resource gap parameter for the i-th user in the grid at time t, representing the shortage of user resource demand; The data transmission rate required by the i-th user within the grid at time t; To prevent zero factor; This is the preset fourth adjustment coefficient; For at any time The grid resource surplus index. 9.The AI-based resource scheduling method of claim 8, wherein, The resource gap parameter is calculated using the following formula: ； in, Let be the resource gap parameter for the i-th user in the grid at time t; max(·) is the maximum value function; The data transmission rate required by the i-th user within the grid at time t; Let be the achievable rate of the i-th user within the grid at time t, and let be the maximum data transmission rate that the network can provide to the i-th user at time t.

10. An artificial intelligence-based resource scheduling apparatus, characterized by, include: The grid data acquisition module is configured to divide the cell coverage area into multiple grids and acquire the grid weight, number of users, average service pressure of users within the grid, and actual grid load of each grid at the current time and at a preset number of historical time steps. The grid weight represents the degree to which the priority of users within the grid amplifies the grid load prediction, and the service pressure represents the degree of tension in user resource demand. The grid load prediction module is configured to output the grid load prediction value at the start of the next scheduling cycle based on the grid weight, number of users, average service pressure of users in the grid and actual grid load of each grid at the current time and a preset number of historical time steps. The grid resource allocation module is configured to allocate the total bandwidth resources of the cell to each grid according to the proportion of the grid priority score of each grid in the total grid priority score of all grids, so as to obtain the amount of schedulable resources for each grid in the current scheduling cycle; wherein the grid priority score is calculated based on the grid load prediction value and the user weight of each user in the grid, and represents the priority of the grid in resource allocation. The user resource scheduling module is configured to allocate the schedulable resources of each grid to each user in the current scheduling cycle according to the proportion of each user's user weight in the total user weight of all users in the grid; wherein the user weight represents the user's priority in resource scheduling.

11. A computer device, comprising: It includes a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the resource scheduling method based on artificial intelligence according to any one of claims 1 to 9.

12. A computer readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by a processor, the processor performs the resource scheduling method based on artificial intelligence according to any one of claims 1 to 9.

Images