Scheduling optimization method and system for network point footprint processing based on genetic algorithm

Abstract

The invention relates to a scheduling optimization method and system for network point footprint processing based on a genetic algorithm, and relates to the technical field of task scheduling. The method comprises the following steps: obtaining network point related data, wherein the network point related data comprises an ID of a network point and AOI area data associated with the network point; obtaining historical execution time consumption of all websites based on the website related data, and generating time sequence data; constructing and calling a hybrid prediction model, and predicting future task execution time consumption of each website by taking the time sequence data as input; and calling a genetic algorithm, and solving and outputting an optimal or approximately optimal task scheduling queue by taking the predicted task execution time consumption as input. The method has the effect of minimizing the overall task completion time.

Description

Technical Field

[0001] This application relates to the technical field of task scheduling, and in particular to a scheduling optimization method and system for processing network footprints based on genetic algorithms. Background Technology

[0002] In backend systems of logistics and supply chain management, batch processing of footprint data (such as GPS trajectory and geofence matching calculations) generated by thousands of physical locations is required daily. Currently, the industry commonly uses task scheduling middleware (such as XXL-Job) to implement distributed processing. Its conventional strategy is to statically distribute the list of locations based on the number of servers, for example, distributing 1,000 locations equally among 10 servers, with each server processing a fixed 100 locations.

[0003] Due to significant differences in business scale and data volume among various network points, the processing time for a single task can range from several seconds to tens of minutes. Therefore, in actual execution, servers assigned to network points with shorter processing times become idle prematurely, while servers assigned to network points with longer processing times run under heavy load for extended periods, becoming bottlenecks in the overall process. This results in uneven utilization of cluster computing resources, low overall system throughput efficiency, and the total task completion time being delayed by the longest task. Summary of the Invention

[0004] In order to minimize the overall task completion time, this application provides a scheduling optimization method and system for network footprint processing based on genetic algorithm.

[0005] On the one hand, the scheduling optimization method for network footprint processing based on genetic algorithm provided in this application adopts the following technical solution: A scheduling optimization method for network footprint processing based on genetic algorithm includes the following steps: Obtain branch-related data, including the branch ID and the AOI area data associated with the branch; Based on the network-related data, the historical execution time of all networks is obtained, and time-series data is generated; A hybrid prediction model is constructed and invoked, using the time-series data as input, to predict the future task execution time of each branch. The genetic algorithm is invoked, taking the predicted task execution time as input, to solve for and output the optimal or near-optimal task scheduling queue.

[0006] By adopting the above technical solution and acquiring network-related data, including network IDs and their associated AOI regions, a basic data framework for the scheduling optimization problem is systematically constructed. Based on this data, time-series data for all networks is generated, comprehensively recording historical task execution. A hybrid prediction model is built and invoked to process this time-series data, enabling proactive prediction of future task execution times for each network, providing crucial input for subsequent optimization. Finally, a genetic algorithm is invoked, using the aforementioned predicted task execution times as optimization costs, to globally search the solution space and output the optimal or near-optimal task scheduling queue. This effectively solves the core scheduling optimization problem of minimizing the total task completion time.

[0007] Preferably, the step of generating the historical time sequence for each network point includes: Using the static framework of network points and AOI areas as an index, and integrating task execution logs and infrastructure resource monitoring data, a historical task performance detail table is generated. Set a time window, and aggregate the data in the historical task performance details table using the time window to obtain the aggregated time consumption index for each time window; The aggregated time consumption index is connected to each network point in chronological order to construct the historical time consumption sequence, and the historical time consumption sequence is time-series data.

[0008] By adopting the above technical solution, a regular historical time series can be constructed, providing a high-quality input data foundation for subsequent prediction models. Preferably, the step of constructing and calling the hybrid prediction model, using the historical time consumption sequence as input, to predict the future task execution time of each branch includes: The historical time-consuming sequence is normalized, and training samples for the hybrid prediction model are constructed using a sliding window method with a preset time step T. Construct a hybrid prediction model that includes a convolutional neural network and a long short-term memory network, and train the hybrid prediction model using the training samples; From the historical time consumption sequence of the network points to be predicted, extract the latest aggregated time consumption index data for T consecutive time windows to form the model input data; Input the model input data into the trained hybrid prediction model and obtain the prediction task time.

[0009] By adopting the above technical solution, the historical time consumption sequence is normalized and training samples are constructed according to a preset time step T. A hybrid prediction model of CNN and LSTM is trained, and the latest T consecutive time window data of the network point to be predicted is extracted and input into the model, thus realizing accurate prediction of the execution time of future tasks.

[0010] Preferably, the training samples for constructing the model include: The time-series data is normalized. Using a preset fixed-length time step, continuous subsequences are extracted from the normalized sequence using a sliding window method, with each subsequence serving as an input feature for a training sample. The data from the time point immediately following each of the subsequences is used as the predicted label for that training sample.

[0011] By adopting the above technical solution, and using a sliding window method with a fixed time step, continuous subsequences are extracted from the normalized time series data as input features, and the data at the next time point is used as the prediction label, structured training samples are provided for the supervised learning model.

[0012] Preferably, the hybrid prediction model comprises, in sequence: An input layer is used to receive the input data and set the time step of the input data to T; A one-dimensional convolutional layer is used to extract local temporal features from the input data; A pooling layer is used to downsample the output of the one-dimensional convolutional layer. The flattening layer is used to convert the output of the pooling layer into a format suitable for LSTM processing; LSTM layer stacking, including multiple LSTM units and batch normalization layers, is used to learn long-term dependencies in time series data; The output layer is used to output the time taken for the prediction task.

[0013] By adopting the above technical solution, local fluctuations and short-term patterns in time-consuming sequences are automatically captured by one-dimensional convolutional layers, long-term trends and periodic dependencies are learned by stacking LSTM layers with batch normalization, and finally predicted values ​​are output through fully connected layers, thus achieving hierarchical, efficient extraction and fusion of complex temporal features.

[0014] Preferably, the LSTM layer stack includes an initial LSTM layer, a first batch of normalized layers, a first LSTM layer, a second LSTM layer, a third LSTM layer, and a second batch of normalized layers, which are connected in sequence.

[0015] By adopting the above technical solution, and using a specific structure of alternating stacking of four LSTM units and two batch normalization layers, the gradient problem is effectively alleviated while the distribution of intermediate features is stabilized by batch normalization, thereby promoting stable training and improved generalization performance of deep networks.

[0016] Preferably, the step of calling the genetic algorithm, using the predicted task execution time as input, to solve for and output the optimal or near-optimal task scheduling queue includes: The computational task of each network point is defined as a task to be scheduled, and the scheduling scheme is mapped to a chromosome with the predicted task time as the cost. Generate an initial population by randomly generating multiple different task sequences to form the initial population; Fitness is evaluated for each individual in the initial population, and the fitness is calculated based on the total predicted time of the scheduling scheme corresponding to that individual. Filter parent sequences; based on the fitness, filter parent sequences from the current population. Generate offspring individuals, and perform a crossover operation on the selected parent individuals to recombine their task sequence fragments to generate offspring sequences; Set the number of iterations and repeat the process of evaluating fitness, selecting parent sequences, and generating offspring individuals until the number of repetitions equals the number of iterations. During the iteration process, execute an elite retention strategy, which includes replacing the worst individual in the offspring population with the best individual in the parent population.

[0017] By adopting the above technical solution, the scheduling scheme is encoded as an evolvable chromosome. Through cyclic optimization using standard evolutionary operators such as initialization, fitness evaluation, selection, crossover, iteration, and elite retention, the algorithm can efficiently search and converge in a discrete permutation and combination solution space in the direction of minimizing the total prediction time.

[0018] Preferably, after performing a crossover operation on the parent individual to generate offspring individuals, a mutation operation is performed on the generated offspring individuals; the mutation operation includes setting a mutation probability and randomly swapping the task codes at two different positions in the task sequence of the offspring individuals according to the mutation probability.

[0019] By adopting the above technical solution, and by randomly exchanging the codes of two positions in the task sequence of the offspring individuals generated through crossover with a certain probability, a controllable random perturbation is introduced, which helps to maintain the genetic diversity of the population, break the local optimum pattern that the algorithm may fall into, and thus enhance the algorithm's exploration ability in the global scope.

[0020] On the other hand, this application also provides a scheduling optimization system for network footprint processing based on genetic algorithms, which adopts the following technical solution: A scheduling optimization system for network footprint processing based on genetic algorithms, comprising the following components: The data acquisition and processing unit is used to acquire relevant data of the network points and generate historical time-series data of each network point. The task time prediction unit is used to construct and call the hybrid prediction model, using the time series data as input, to predict the future task execution time of each branch. The task scheduling optimization unit is used to call the genetic algorithm, taking the predicted task execution time as input, to solve and output the optimal or near-optimal task scheduling queue.

[0021] By adopting the above technical solution, the responsibilities and data interfaces of the three major functional modules of data processing, time prediction and scheduling optimization are clarified through modular design, realizing the modular, integrated and automated operation of the technical solution, so that the method can play a role as a deployable service or product.

[0022] Preferably, the task scheduling optimization unit includes: The encoding module is used to encode the task scheduling scheme into chromosomes; Genetic algorithm engine, used to manage the population and generate optimal or near-optimal chromosomes; The decoding module is used to decode the optimal or near-optimal chromosome into a list of task execution orders.

[0023] By adopting the above technical solution, the business problem is transformed into an algorithmic problem through the encoding module, optimization calculations are performed through the genetic algorithm engine, and the optimization results are restored to business instructions through the decoding module. This internal structure design clearly separates the logic of problem modeling, solution calculation, and result parsing, improving the independence, reusability, and maintainability of the unit and the entire system.

[0024] In summary, this application includes at least one of the following beneficial technical effects: 1. By constructing historical time consumption sequences and inputting them into a hybrid prediction model to predict future task time consumption, and then using the predicted time consumption as the cost input to a genetic algorithm to solve for the optimal task scheduling queue, global optimization of task scheduling based on accurate forward prediction is achieved, which effectively improves the overall processing efficiency of the system. 2. A hybrid prediction model consisting of a one-dimensional convolutional neural network and a long short-term memory network is used to process time-series data. Local time-series features are extracted using one-dimensional convolutional layers, and long-term dependencies are learned by stacking LSTM layers, which significantly improves the prediction accuracy of future changes in the execution time of network tasks. 3. By encoding the scheduling scheme into chromosomes and using the selection, crossover, mutation, and elite retention mechanisms of genetic algorithms for iterative optimization, it is possible to efficiently search and approximate the task execution sequence with the shortest total time within a huge solution space, thus solving the scheduling and sorting problem in complex scenarios. Attached Figure Description

[0025] Figure 1 This is an overall flowchart of Embodiment 1 of this application; Figure 2 This is a structural diagram of the hybrid prediction module in Embodiment 1 of this application; Figure 3 This is a flowchart of step S4 in Embodiment 1 of this application; Figure 4 This is a schematic diagram of Embodiment 2 of this application. Detailed Implementation

[0026] The following combination Figures 1-4 This application will be described in further detail. Example

[0027] Embodiment 1 of this application discloses a scheduling optimization method for network footprint processing based on genetic algorithm.

[0028] Reference Figure 1 A scheduling optimization method for network footprint processing based on genetic algorithms includes the following steps: S1: Obtain relevant data about the outlets; S2: Based on the network location-related data, obtain the historical execution time of all network locations and generate time-series data; S3: Build and call the hybrid prediction model, using time series data as input, to predict the future task execution time of each branch; S4: Call the genetic algorithm, using the predicted task execution time as input, to solve for and output the optimal or near-optimal task scheduling queue.

[0029] Specifically, step S1 involves acquiring relevant data about the network outlets, including the following: By directly querying the internal business database, a list of static basic information for all active outlets can be obtained. The list of static basic information includes fields such as the ID, physical address, and operational status of multiple outlets.

[0030] To determine the service area of ​​each location, the system needs to further acquire the AOI (Area of ​​Interest) data associated with each location. AOI data typically originates from a professional geographic information system, obtained by calling an internal GIS database or an API from a third-party map service provider. An AOI data package usually includes a unique identifier and name for each AOI, as well as closed polygon boundary data consisting of a series of latitude and longitude coordinates.

[0031] The system also needs to process the continuously generated, massive amounts of real-time GPS location data. This real-time GPS location data is primarily obtained by subscribing to a message middleware, with the data source being courier applications or logistics devices; it may also be supplemented by periodically fetching batch data files stored on a distributed file system.

[0032] Step S2, based on the network location-related data, obtains the historical execution time of all network locations and generates time-series data, including the following: Using the site ID and AOI area data output in step S1 as indexes, deep fusion and processing are performed on various raw dynamic data streams from the computing cluster. The raw dynamic data mainly includes two categories: first, task execution logs that record the details of each GPS-AOI matching calculation event, including the task's start time, end time, execution status, and the identifier of the server to which it was scheduled; second, snapshots of infrastructure resources periodically collected by the monitoring system, which record performance indicators such as CPU utilization and memory usage of each server at specific points in time, thereby generating a historical task performance detail table, including dimensions such as task number, task duration, task association, and machine resources.

[0033] The task number is generated by assigning a globally unique identifier to each task log entry, and a globally unique ID is generated for each computation task to ensure the independence and traceability of each task. The task duration is derived by calculating the difference between the start time and end time in the task log; the task association is determined by associating and matching the task log with the static framework of S1 to clarify the network site and AOI area to which the task belongs; machine resources are obtained by associating and aggregating (e.g., calculating the average value) the server identifier in the task log with the resource monitoring snapshots of the corresponding time period.

[0034] Finally, all task records for each site are aggregated within a fixed time window (e.g., 5 minutes). The total and average time of all tasks within the time window are calculated to obtain the aggregated time for that time window, and a timestamp for that time window is generated. Subsequently, the aggregated time data points generated by each site, arranged in chronological order, are sequentially connected to construct a complete and continuous historical time sequence for that site, and this historical time sequence is output as time-series data to subsequent steps.

[0035] S3: Build and invoke a hybrid forecasting model, using time-series data as input, to predict the future task execution time for each branch, including the following: S31: Construct training samples for the hybrid prediction model.

[0036] The time-series data output from step S2 is received and normalized to eliminate the influence of dimensions. Then, the time step needs to be pre-set according to the prediction task's objective and model design. In this embodiment, the object whose future task duration needs to be predicted is the network point to be predicted. When making predictions, the input data used is the most recent continuous historical time-series data that meets the model's time step requirements and is available at the prediction time for the network point to be predicted. In this embodiment, the time step is set to 45, meaning that the model will review and base its predictions on the time-series data of the most recent 45 consecutive time windows.

[0037] Subsequently, according to the set time step, the long sequence is divided into multiple time-series training samples using the sliding window method. Each sample consists of the aggregated time consumption of 45 consecutive time windows, which serves as the input feature, and the aggregated time consumption of the next time window immediately following these 45 windows serves as the prediction label, thus providing the target value required for training the subsequent supervised learning hybrid prediction model.

[0038] S32: Construct a hybrid prediction model that includes a convolutional neural network and a long short-term memory network. Train the hybrid prediction model using training samples, input the latest time series data of the points to be predicted into the trained model, and obtain the prediction task time output by the hybrid prediction model.

[0039] Reference Figure 2 The hybrid prediction model includes an input layer, a one-dimensional convolutional layer, a pooling layer, a flattening layer, a stack of LSTM layers, a random deactivation layer, and an output layer.

[0040] The input layer receives multiple training samples. In this embodiment, the shape of the training samples is defined as (None, 45, 1), where None represents the batch size, which is dynamically determined by the actual number of samples input during training or prediction, ensuring the flexibility of the model; 45 represents the pre-set time step; and 1 represents the feature dimension. In this embodiment, only one core feature is used in each time window, i.e., the aggregation time of that window. The function of the input layer is to receive and verify whether the input data conforms to the shape specification of (None, 45, 1), and output the input data to the next layer as is. Therefore, the input and output of the input layer are consistent in shape, both being [(None, 45, 1)].

[0041] One-dimensional convolutional layers are used for feature extraction, learning local features from the temporal training samples output by the input layer. By using multiple independent filters, the one-dimensional convolutional layer performs one-dimensional convolution operations on the received data, aiming to automatically extract and enhance short-term fluctuations and local dependency patterns from 45 consecutive time windows. In this embodiment, the convolutional kernel is set to 2, meaning that each filter simultaneously observes two adjacent time windows in the temporal training samples during each computation, thereby capturing the local relationships between them. In this embodiment, the output shape of the one-dimensional convolutional layer is [(None, 44, 128)]. Here, None indicates that the batch size remains unchanged and consistent with the input layer; 44 indicates that the time step changes from 45 to 44, because a convolutional kernel of length 2 sliding over a time step of length 45 will generate 44 effective convolutional positions; 128 indicates that the feature dimension changes from 1 to 128, generating 128 different feature maps, each of which extracts a specific local pattern from the original time-consuming sequence. One-dimensional convolutional layers can transform a single time-consuming sequence into a new, shorter sequence containing 128 rich local features, thereby automatically extracting and enhancing short-term fluctuations and local dependency patterns in continuous time-series training samples.

[0042] Pooling layers are used to downsample the feature sequences output by one-dimensional convolutional layers. By aggregating local features, they retain the most salient information while compressing computation and improving the model's robustness to small perturbations. In this embodiment, the pooling window `pool_size` is set to 1, meaning the pooling window only covers a single time step, thus not changing the length and dimension of the time step. After processing by the pooling layer, the output shape of the pooling layer remains [(None, 44, 128)].

[0043] The flattening layer is used to convert the multi-dimensional feature map at each time step of the pooling layer output into a one-dimensional feature vector format suitable for LSTM processing. The output shape of the flattening layer is [(None, 44, 128)], which means that the output of the flattening layer contains 44 time steps, each time step is represented by a 128-dimensional vector, so that subsequent LSTM layers can process it according to time steps.

[0044] The LSTM layer stack consists of an initial LSTM layer, a first batch of normalization layers, a first LSTM layer, a second LSTM layer, a third LSTM layer, and a second batch of normalization layers, all connected in sequence. In this embodiment, the initial LSTM layer has 50 units. It receives the output from the flattening layer, and the output shape of the first LSTM layer is [(None, 44, 50)], indicating that the first LSTM layer transforms and condenses the 128-dimensional input features at each time step into a 50-dimensional state vector. The state vector encodes all temporal patterns and contextual information learned by the model from historical time-consuming data up to the current time step. The first batch of normalization layers is used to standardize the state vector output by the initial LSTM layer. In this embodiment, the state vector is adjusted to a distribution with a mean of 0 and a variance of 1, and the output shape of the first batch of normalization layers is [(None, 44, 50)].

[0045] The first LSTM layer further deepens the learning of temporal dependencies within the same feature space using the batch-normalized state vector. The output state of the first LSTM layer remains [(None, 44, 50)]. The second LSTM layer maps the learned patterns to a higher-dimensional feature space, and its output state is [(None, 44, 128)]. The feature dimension expands from 50 to 128, indicating that the expressive power of the hidden state is enhanced to 128 dimensions at a deeper level to encode more complex temporal patterns. The third LSTM layer receives the output of the second LSTM layer, returning only the state vector of the last time step, and its output shape is [(None, 50)]. The sequence length dimension is removed, and the feature dimension is condensed from 128 to 50. The output 50-dimensional vector is the predicted feature vector learned by the model from the entire 44-step, 128-dimensional high-dimensional sequence information. The second batch of normalization layers normalizes the predicted feature vector output by the third LSTM layer, providing a stable input distribution for the final linear mapping, and the output shape of the second batch of normalization layers is [(None, 50)].

[0046] The random deactivation layer introduces a regularization effect by randomly discarding some neuron connections during forward propagation, aiming to prevent the model from overfitting the training data. The output shape of the random deactivation layer is [(None, 50)]. The output layer is the decision endpoint and business interface of the model. It receives the 50-dimensional feature vector processed and refined by all the previous layers and maps the multi-dimensional adjustment vector to a single scalar value [(None, 1)] through linear transformation. The scalar value represents the final predicted task time of the model for the next time point of the target network node.

[0047] Reference Figure 3S4: Call the genetic algorithm, using the predicted task execution time as input, to solve for and output the optimal or near-optimal task scheduling queue, including the following steps: S41: Problem Modeling and Coding.

[0048] First, the task scheduling problem is formalized. The GPS-AOI matching calculation for each network point is defined as a task. Each task is associated with two key attributes: task ID (unique identifier) ​​and predicted task time. The scheduling resource is a single computation queue, meaning only one task is executed at any given time, and there is no priority preemption between tasks. The optimization objective of step S41 is to find a task execution order that minimizes the total time required to complete all tasks. This embodiment uses direct sequence encoding, where a chromosome represents a complete task scheduling scheme, and the gene sequence represents the task execution order. For example, the chromosome [network point A, network point C, network point B, ...] indicates that all calculations for network point A are executed first, followed by network point C, and so on. Each different permutation represents a scheduling scheme, constituting the solution space that the algorithm needs to search.

[0049] S42: Generate the initial population and generate multiple task scheduling schemes.

[0050] An initial population size is set, and different task sequences for each network point are randomly generated based on this initial population size as a task scheduling scheme. For example, the initial population size can be set to 100. The generation of each network point's task sequence is achieved by uniformly and randomly shuffling all late-delay task lists, thereby ensuring a uniform distribution of the initial population in the solution space.

[0051] S43: Calculate the fitness of each task scheduling scheme.

[0052] A fitness function is set up, where fitness is equal to the reciprocal of the total task scheduling time. For a single task, the total scheduling time is obtained by summing the predicted task times of the corresponding nodes according to the task sequence. The shorter the total task scheduling time, the higher the fitness of the task. This allows the evaluation of the task scheduling scheme to be based on the prediction of future performance, rather than historical data, thus achieving forward-looking scheduling.

[0053] S44: Select parent sequences based on fitness.

[0054] Simulating natural selection, this algorithm probabilistically selects task scheduling sequences for reproducing the next generation based on fitness levels. This embodiment employs a roulette wheel selection method. First, the sum of the fitness levels of all individuals in the population is calculated. Then, the selection probability of each individual is calculated, where the selection probability of each individual is equal to the ratio of its fitness to the sum of the fitness levels. This is equivalent to creating a roulette wheel, where the area occupied by each individual on the wheel is proportional to its fitness. The algorithm simulates rotating this roulette wheel, setting the number of selections to 6 (or depending on the number of parent sequences required), and then randomly selecting an individual each time as the parent sequence. Individuals with higher fitness have a significantly greater probability of being selected. This ensures that more efficient schemes have a higher chance of being retained and passed on to the next generation, thus ensuring that the search direction is towards a better scheduling scheme.

[0055] S45: Combine fragments of different parent sequences to generate child sequences.

[0056] Set the crossover probability (e.g., 0.65) and perform gene recombination on the selected parent individuals using the preset crossover probability. Specifically, adopt methods suitable for permutation coding, such as sequential crossover, to randomly select two crossover points. The task sequence fragment of parent sequence one between the two crossover points is completely preserved and extended to the corresponding position in the offspring sequence. Then, remove the task contained in the fragment from parent sequence two and fill the vacant positions in the offspring sequence with the remaining tasks in sequence, thereby generating multiple offspring sequences.

[0057] S46: Perform mutation operation on the combined offspring sequence.

[0058] By introducing new mutant genes into the population through small-scale random changes, a mutation probability (e.g., 0.15) is set, and an individual is determined to mutate based on this probability. If a mutation occurs, two different positions in the offspring sequence are randomly selected for task encoding swapping, thereby generating an updated offspring sequence. When the population tends to be consistent (e.g., all adopting a similar "short task first" arrangement), mutation can disrupt this pattern, prioritizing a longer task to explore neglected solution space regions. Even near a good solution, swapping adjacent tasks may reveal a better local arrangement. The low mutation probability ensures that good patterns are preserved with high probability, with only tentative changes in a few cases.

[0059] S47: Implement an elite retention strategy.

[0060] After crossover and mutation are completed to generate offspring, the individual with the highest fitness in the parent population (i.e., the currently found optimal task scheduling scheme) is recorded. Then, the individual with the lowest fitness in the offspring population (the worst new scheme) is identified, and the elite individual from the parent generation directly replaces the worst individual from the offspring. This ensures that the historically optimal task scheduling scheme discovered during the evolutionary process is never lost due to the randomness of selection, crossover, and mutation.

[0061] S48: Perform population iteration.

[0062] The number of iterations is set; in this embodiment, it is set to 90. Steps S43 to S47 are repeated, with each completed loop counting as one iteration, until the set number of iterations is reached, at which point the iteration stops. As the number of iterations accumulates, the genetic patterns of individuals in the population are continuously recombined and screened through selection, crossover, and mutation. This process drives a stepwise increase in the overall fitness of the population, characterized by a gradual reduction in the total time taken by the discovered task scheduling scheme. The iterative mechanism essentially simulates multiple rounds of generation, testing, and selection optimization cycles, allowing the scheduling scheme to accumulate and improve across generations, gradually evolving from an initial random arrangement into a highly specialized task execution order, ultimately achieving the core objective of minimizing the overall completion time.

[0063] S49: After the iteration is completed, select the individual with the highest fitness from the final population, and generate a task execution order list by decoding the individual with the highest fitness.

[0064] The individual with the highest fitness represents the optimal or near-optimal task scheduling scheme obtained after multiple generations of iterative evolution. Since this application uses direct sequence encoding, the gene sequences on the chromosomes are read sequentially during decoding. For example, if the chromosome of the individual with the highest fitness is [D, A, C, B], then the decoding result is the task execution order list: node D -> node A -> node C -> node B. The decoded task order is used as the final output and can be directly delivered to the task dispatcher for execution.

[0065] The implementation principle of the scheduling optimization method for network footprint processing based on genetic algorithm in Embodiment 1 of this application is as follows: It integrates steps such as data acquisition, prediction model construction, and genetic algorithm optimization, and fully considers the actual situation of the network and the dynamic changes of the task. By accurately acquiring relevant network data and historical execution time, it provides rich information for the prediction model, making the prediction of task time more accurate. The introduction of genetic algorithm, through operations such as chromosome encoding, fitness calculation, selection, crossover, mutation, and elite retention, continuously searches for the optimal task scheduling scheme in the solution space, avoiding the subjectivity and rigidity of traditional scheduling methods, improving the efficiency and accuracy of scheduling, and minimizing the total task time. It has made significant improvements and contributions to the existing technology in network task scheduling. Example

[0066] Embodiment 2 of this application discloses a scheduling optimization system for network footprint processing based on genetic algorithm.

[0067] refer to Figure 4 A scheduling optimization system for network footprint processing based on genetic algorithms includes a data acquisition and processing unit, a task time prediction unit, a task scheduling optimization unit, and a task execution module.

[0068] The data acquisition and processing unit is responsible for collecting and processing data related to service points. It obtains service point and AOI data by querying business databases and geographic information systems, and connects to real-time GPS data streams through message middleware. Its core function is to integrate task logs and resource monitoring data from the computing cluster to generate a historical task performance detail table containing task number, duration, relationships, and machine resources. This table is further aggregated into a regularized historical duration sequence for each service point and output to the task duration prediction unit.

[0069] The task time prediction unit receives historical time sequences and outputs future time predictions. This module incorporates a pre-trained CNN-LSTM hybrid prediction model. The hybrid prediction model first normalizes the input sequence and constructs samples at fixed time steps. Then, it extracts local features through a one-dimensional convolutional layer, learns long-term temporal dependencies through a multi-layer long short-term memory network, and finally outputs the predicted task time for the next time window corresponding to the network point through a fully connected layer. The prediction results are then sent to the task scheduling and optimization unit.

[0070] The task scheduling and optimization unit first defines each network task as a task to be scheduled, and encodes the task execution order as a chromosome, using the predicted execution time as the cost. The genetic algorithm engine manages the population, calculates the fitness based on the reciprocal of the total predicted execution time, and iteratively executes evolutionary operations such as roulette wheel selection, sequential crossover, and exchange mutation. Simultaneously, an elite retention strategy is employed to ensure that optimal or near-optimal solutions are not lost. After the algorithm reaches a specified number of generations, the decoding module converts the obtained optimal or near-optimal chromosomes into a specific list of task execution orders, which is then output to the task execution module.

[0071] The task execution module receives the optimal scheduling queue and strictly follows this order to drive the servers or containers in the backend computing resource pool to execute the actual GPS-AOI matching calculation tasks, thereby completing all calculation work with globally optimal efficiency and maximizing system throughput.

[0072] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A scheduling optimization method for network footprint processing based on genetic algorithm, characterized in that: Includes the following steps: Obtain branch-related data, including the branch ID and the AOI area data associated with the branch; Based on the network-related data, the historical execution time of all networks is obtained, and time-series data is generated; A hybrid prediction model is constructed and invoked, using the time-series data as input, to predict the future task execution time of each branch. The genetic algorithm is invoked, taking the predicted task execution time as input, to solve for and output the optimal or near-optimal task scheduling queue.

2. The scheduling optimization method for network footprint processing based on genetic algorithm according to claim 1, characterized in that: The step of generating the historical time sequence for each network point includes: Using the static framework of network points and AOI areas as an index, and integrating task execution logs and infrastructure resource monitoring data, a historical task performance detail table is generated. Set a time window, and aggregate the data in the historical task performance details table using the time window to obtain the aggregated time consumption index for each time window; The aggregated time consumption index is connected to each network point in chronological order to construct the historical time consumption sequence, and the historical time consumption sequence is time-series data.

3. The scheduling optimization method for network footprint processing based on genetic algorithm according to claim 1, characterized in that: The steps of constructing and calling the hybrid prediction model, using the historical time consumption sequence as input, to predict the future task execution time of each branch include: The historical time-consuming sequence is normalized, and training samples for the hybrid prediction model are constructed using a sliding window method with a preset time step T. Construct a hybrid prediction model that includes a convolutional neural network and a long short-term memory network, and train the hybrid prediction model using the training samples; From the historical time consumption sequence of the network points to be predicted, extract the latest aggregated time consumption index data for T consecutive time windows to form the input data of the model; Input the model input data into the trained hybrid prediction model and obtain the prediction task time.

4. The scheduling optimization method for network footprint processing based on genetic algorithm according to claim 3, characterized in that: The training samples for constructing the model include: The time-series data is normalized. Using a preset fixed-length time step, continuous subsequences are extracted from the normalized sequence using a sliding window method, with each subsequence serving as an input feature for a training sample. The data from the time point immediately following each of the subsequences is used as the predicted label for that training sample.

5. The scheduling optimization method for network footprint processing based on genetic algorithm according to claim 3, characterized in that: The hybrid prediction model includes, in sequence: An input layer is used to receive the input data and set the time step of the input data to T; A one-dimensional convolutional layer is used to extract local temporal features from the input data; A pooling layer is used to downsample the output of the one-dimensional convolutional layer. The flattening layer is used to convert the output of the pooling layer into a format suitable for LSTM processing; LSTM layer stacking, including multiple LSTM units and batch normalization layers, is used to learn long-term dependencies in time series data; The output layer is used to output the time taken for the prediction task.

6. The scheduling optimization method for network footprint processing based on genetic algorithm according to claim 5, characterized in that: The LSTM layer stack includes an initial LSTM layer, a first batch of normalized layers, a first LSTM layer, a second LSTM layer, a third LSTM layer, and a second batch of normalized layers, which are connected by signals in sequence.

7. The scheduling optimization method for network footprint processing based on genetic algorithm according to claim 1, characterized in that: The step of calling the genetic algorithm, using the predicted task execution time as input, to solve for and output the optimal or near-optimal task scheduling queue includes: The computational task of each network point is defined as a task to be scheduled, and the scheduling scheme is mapped to a chromosome with the predicted task time as the cost. Generate an initial population by randomly generating multiple different task sequences to form the initial population; Fitness is evaluated for each individual in the initial population, and the fitness is calculated based on the total predicted time of the scheduling scheme corresponding to that individual. Filter parent sequences; based on the fitness, filter parent sequences from the current population. Generate offspring individuals, and perform a crossover operation on the selected parent individuals to recombine their task sequence fragments to generate offspring sequences; Set the number of iterations and repeat the process of evaluating fitness, selecting parent sequences, and generating offspring individuals until the number of repetitions equals the number of iterations. During the iteration process, execute an elite retention strategy, which includes replacing the worst individual in the offspring population with the best individual in the parent population.

8. The scheduling optimization method for network footprint processing based on genetic algorithm according to claim 7, characterized in that: After performing a crossover operation on the parent individual to generate offspring individuals, a mutation operation is performed on the generated offspring individuals; the mutation operation includes setting a mutation probability and randomly swapping the task codes at two different positions in the task sequence of the offspring individuals according to the mutation probability.

9. A scheduling optimization system for network footprint processing based on genetic algorithm, wherein the method described in any one of claims 1-8 is characterized in that, include: The data acquisition and processing unit is used to acquire relevant data of the network points and generate historical time-series data of each network point. The task time prediction unit is used to construct and call the hybrid prediction model, using the time series data as input, to predict the future task execution time of each branch. The task scheduling optimization unit is used to call the genetic algorithm, taking the predicted task execution time as input, to solve for and output the optimal or near-optimal task scheduling queue.

10. A scheduling optimization system for network footprint processing based on genetic algorithm according to claim 9, characterized in that: The task scheduling optimization unit includes: The encoding module is used to encode the task scheduling scheme into chromosomes; Genetic algorithm engine, used to manage the population and generate optimal or near-optimal chromosomes; The decoding module is used to decode the optimal or near-optimal chromosome into a list of task execution orders.

Images