Computing power demand prediction and resource pre-distribution system based on deep learning

Abstract

The invention discloses a computing power demand prediction and resource pre-distribution system based on deep learning, and relates to the technical field of artificial intelligence and computing resource scheduling management, and the system comprises a data collection fusion module, a demand prediction and correction module, a strategy generation and verification module, and a model updating module. And the data acquisition and fusion module is used for acquiring multi-source heterogeneous data and carrying out preprocessing and feature fusion on the data. According to the computing power demand prediction and resource pre-distribution system based on deep learning, a prediction process containing multi-source data fusion and having a dual detection and correction mechanism is constructed, so that the dynamics and uncertainty of the computing power demand are effectively coped with; by means of the design, the prediction result can still keep high credibility in the data exception or burst mode, a stable and reliable data basis is provided for subsequent resource decision making, and therefore the risk of scheduling errors caused by prediction deviation is reduced from the source.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and computing resource scheduling and management technology, specifically to a deep learning-based computing power demand prediction and resource pre-allocation system. Background Technology

[0002] With the widespread adoption of cloud computing, edge computing, and artificial intelligence, the efficient scheduling and management of computing resources has become a critical issue for data centers, cloud service platforms, and intelligent computing systems. Computing demands are characterized by dynamism, uncertainty, and suddenness. Traditional resource allocation methods, typically based on static rules, fixed thresholds, or simple historical averages for prediction and scheduling, are ill-suited to complex and ever-changing real-world application scenarios. These methods often lead to uneven resource utilization, increased response latency, higher energy consumption, and even system performance bottlenecks due to resource contention.

[0003] In existing technologies, there are several resource scheduling schemes based on rules or traditional machine learning. The patent publication "CN116542491A" describes "An Artificial Intelligence Task Scheduling Method and System Based on Deep Learning." This system collects historical data and road condition information of repair personnel, combines it with a deep learning model to generate a scheduling index, and achieves intelligent scheduling of repair personnel. This technology improves the accuracy and efficiency of task scheduling to a certain extent, but still has the following shortcomings: First, its scheduling logic mainly revolves around fixed dimensions such as personnel capabilities and road conditions, without fully considering the dynamic changes in computing power demand itself over time, task type, and system state; second, its model relies on the completeness of historical data, making it less adaptable to scenarios with sudden, high-concurrency computing power demands; third, the system does not achieve proactive pre-allocation of resources, remaining a passive response scheduling method, making it difficult to prepare resources before peak demand periods, easily leading to resource shortages or idleness.

[0004] Therefore, existing technologies still suffer from problems such as inaccurate prediction, scheduling lag, and low resource utilization when dealing with dynamic, high-concurrency, and multi-type computing power demands, and cannot achieve truly intelligent and forward-looking resource management. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning-based computing power demand prediction and resource pre-allocation system to solve the problems mentioned in the background art.

[0006] To solve the above technical problems, the present invention provides the following technical solution: a deep learning-based computing power demand prediction and resource pre-allocation system, including a data acquisition and fusion module, a demand prediction and correction module, a strategy generation and verification module, and a model update module; The data acquisition and fusion module is used to acquire multi-source heterogeneous data and perform preprocessing and feature fusion on the data; The demand forecasting and correction module is connected to the data acquisition and fusion module. It is used to generate a preliminary computing power demand forecasting sequence based on the fused data, perform anomaly detection and correction on the preliminary computing power demand forecasting sequence, and output the corrected computing power demand sequence. The strategy generation and verification module is connected to the demand prediction and correction module, and is used to generate a resource pre-allocation strategy based on the corrected computing power demand sequence, and to verify and evaluate the resource pre-allocation strategy in a simulation environment. The model update module is connected to the data acquisition and fusion module, the demand prediction and correction module, and the strategy generation and verification module, respectively. It is used to collect actual system operation data and perform online incremental learning updates on the deep learning models involved in the system based on the actual operation data.

[0007] Furthermore, the demand forecasting and correction module includes a forecasting unit, a first detection unit, and a correction unit; The prediction unit is connected to the data acquisition and fusion module. It has a multi-head spatiotemporal attention neural network model built inside it to receive the fused data and output a preliminary computing power demand prediction sequence. The first detection unit is connected to the prediction unit and is used to perform real-time abnormal pattern recognition on the preliminary computing power demand prediction sequence based on an adaptive threshold algorithm. The abnormal patterns include numerical spikes, trend drifts, or periodic breaks. The correction unit is connected to the first detection unit and the prediction unit respectively. When the first detection unit identifies an anomaly, the correction unit is activated. It has a pre-set generative adversarial network model. The generator of the generative adversarial network model is used to reconstruct the identified abnormal data fragments to replace the original abnormal fragments and generate an intermediate correction sequence.

[0008] Furthermore, the demand forecasting and correction module also includes a second detection unit; The second detection unit is connected to the correction unit and is used to perform anomaly detection based on the adaptive threshold algorithm on the intermediate correction sequence again; If the second detection unit does not detect any abnormality, the intermediate correction sequence is output as the final corrected computing power requirement sequence to the strategy generation and verification module. If the second detection unit detects an anomaly, it triggers the correction unit to reconstruct the new abnormal segment, or triggers a manual intervention process.

[0009] Furthermore, the strategy generation and verification module includes a strategy generation unit and a simulation verification unit; The strategy generation unit is connected to the demand prediction and correction module. It has a deep reinforcement learning model built inside. The deep reinforcement learning model takes the corrected computing power demand sequence as input and outputs a resource pre-allocation strategy that includes resource reservation, allocation ratio and scheduling time sequence. The simulation verification unit is connected to the strategy generation unit. It has a lightweight system simulation environment built inside to simulate the execution of the resource pre-allocation strategy. Based on the load balancing index, task completion time index and resource utilization index generated during the simulation execution, the comprehensive evaluation score of the resource pre-allocation strategy is calculated.

[0010] Furthermore, the simulation verification unit and the strategy generation unit form a closed loop; If the overall evaluation score is higher than the preset qualification threshold, the resource pre-allocation strategy is allowed to be deployed and executed in the real system. If the comprehensive evaluation score is lower than the qualified threshold, the evaluation result and the corresponding scenario data are fed back to the policy generation unit, triggering the deep reinforcement learning model to regenerate a new resource pre-allocation policy, or triggering manual policy adjustment.

[0011] Furthermore, the model update module includes a data collection unit and an incremental learning unit; The data collection unit is used to collect actual task execution time data, actual resource consumption data, and task execution status data after the resource pre-allocation strategy verified by the strategy generation and verification module is executed in the real system. The incremental learning unit is connected to the data collection unit and is used to use the actual task execution time data, real resource consumption data, and task execution status data as new training samples to update and optimize the parameters of the neural network model in the demand prediction and correction module and the deep reinforcement learning model in the policy generation and verification module without restarting the system.

[0012] Furthermore, the multi-source heterogeneous data includes task attribute data, system status data, external environment data, and real-time event stream data; The task attribute data includes task type, task priority, inter-task dependencies, and historical execution time records; The system status data includes processor utilization data, memory usage data, network bandwidth usage data, and storage input / output data; The external environment data includes user access pattern data and business cycle tag data; The real-time event stream data includes task submission queue data, system anomaly alarm log data, and resource release signal data.

[0013] A deep learning-based method for predicting computing power demand and pre-allocating resources, applied to the aforementioned system, includes the following steps: S1: Collect multi-source heterogeneous data through the data acquisition and fusion module and perform preprocessing and feature fusion; S2: The demand forecasting and correction module generates a preliminary computing power demand forecasting sequence based on the fused data, and performs at least one anomaly detection and reconstruction correction on the preliminary computing power demand forecasting sequence to obtain the corrected computing power demand sequence. S3: The strategy generation and verification module generates a resource pre-allocation strategy based on the modified computing power demand sequence, and verifies and evaluates the resource pre-allocation strategy in a simulation environment; S4: Deploy and execute the resource pre-allocation strategy that has passed the verification and evaluation in the real system; S5: Collects actual system operation data through the model update module and performs online incremental learning updates on the deep learning models involved in the system.

[0014] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the deep learning-based method for predicting computing power demand and pre-allocating resources.

[0015] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the deep learning-based method for predicting computing power demand and pre-allocating resources.

[0016] This invention provides a deep learning-based system for predicting computing power demand and pre-allocating resources. It offers the following advantages: This deep learning-based computing power demand prediction and resource pre-allocation system effectively addresses the dynamic and uncertain nature of computing power demand by constructing a prediction process that includes multi-source data fusion and dual detection and correction mechanisms. This design ensures that the prediction results maintain high reliability even when facing data anomalies or sudden patterns, providing a stable and reliable data foundation for subsequent resource decisions, thereby reducing the risk of scheduling errors caused by prediction bias from the source.

[0017] This deep learning-based computing power demand prediction and resource pre-allocation system, through the integration of strategy simulation verification and online incremental learning mechanisms, realizes the transformation of resource allocation from passive response to proactive pre-planning. This method can evaluate and optimize scheduling strategies before actual deployment, avoid potential bottlenecks in advance, and enable the system to continuously self-adjust based on actual operation feedback. This not only improves the efficiency and balance of cluster resource utilization, but also enhances the system's adaptability to the evolution of long-term task patterns and short-term emergencies. Attached Figure Description

[0018] Figure 1 This is a data flow diagram between modules of a deep learning-based computing power demand prediction and resource pre-allocation system according to the present invention. Figure 2 This is a closed-loop optimization diagram of the strategy generation and verification module of a deep learning-based computing power demand prediction and resource pre-allocation system according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 and Figure 2 The present invention provides a technical solution: a computing power demand prediction and resource pre-allocation system based on deep learning, including a data acquisition and fusion module, a demand prediction and correction module, a strategy generation and verification module, and a model update module; The data acquisition and fusion module is used to collect multi-source heterogeneous data and perform preprocessing and feature fusion on the data; The demand forecasting and correction module is connected to the data acquisition and fusion module. It is used to generate a preliminary computing power demand forecast sequence based on the fused data, perform anomaly detection and correction on the preliminary computing power demand forecast sequence, and output the corrected computing power demand sequence. The strategy generation and verification module is connected to the demand prediction and correction module. It is used to generate resource pre-allocation strategies based on the corrected computing power demand sequence, and to verify and evaluate the resource pre-allocation strategies in the simulation environment. The model update module is connected to the data acquisition and fusion module, the demand prediction and correction module, and the strategy generation and verification module, respectively. It is used to collect actual system operation data and perform online incremental learning updates on the deep learning models involved in the system based on the actual operation data.

[0021] It should be further explained that after deployment, the system's data acquisition and fusion module continuously acquires data from multiple sources. These sources include, but are not limited to, application programming interfaces (APIs) for submitting tasks, cluster monitoring agents, external calendar services, and the system event bus. This module is configured with dedicated data cleaning and standardization subroutines to handle missing values, anomalous formats, and timestamp alignment in the raw data. After preprocessing, a feature engineering subroutine fuses different types of features; for example, it concatenates periodic external environment labels with real-time system status indicators within the same time window to form a unified multi-dimensional time-series feature vector, providing input for subsequent predictions.

[0022] Subsequently, the demand forecasting and correction module begins operation. At its core is a pre-trained multi-head spatiotemporal attention neural network model. This model receives the aforementioned feature vectors, and its multi-head attention mechanism allows it to capture long-term dependencies and short-term correlations in the data from multiple dimensions, including task history, system load fluctuations, and external events, in parallel, thus outputting a preliminary predicted sequence of computing power demand over a future period. To ensure the reliability of the prediction, this module incorporates a closed-loop verification and correction process. First, a detector based on the Isolation Forest algorithm, coupled with a dynamic threshold adjustment unit, scans the preliminary predicted sequence, identifying potential anomalies that significantly deviate from historical patterns or trends of adjacent points. Once an anomaly is detected, a pre-trained generative adversarial network (GAN) model is activated. The generator of this GAN receives the anomaly and its contextual information, and learns to generate a sequence that conforms to the overall data distribution and has a smooth transition, replacing the marked anomaly segment and completing the first correction. Afterward, the corrected sequence undergoes the same detection logic for secondary verification to ensure the effectiveness of the correction, forming a closed loop of "prediction-detection-correction-re-verification," ultimately outputting a purified and corrected computing power demand sequence.

[0023] Next, the strategy generation and verification module takes action based on the revised demand sequence. This module deploys a deep reinforcement learning agent whose state space is designed to include current and predicted resource requirements, real-time available resources in the cluster, and a queue of tasks to be scheduled. Its action space corresponds to different resource pre-allocation schemes. The agent learns and generates a resource pre-allocation strategy that balances efficiency and stability through interaction with the environment. This strategy specifies the resource reservation amounts on different computing nodes, the allocation priorities for different types of tasks, and the expected scheduling times. Before the strategy is executed in the real environment, it is subjected to a sandbox simulation in a lightweight system simulation environment. This simulation environment simulates scenarios such as the resource scheduler of a real cluster, network latency, and potential node failures. After the simulation, the simulator comprehensively calculates multiple indicators of the strategy during the simulated execution process and provides a quantitative evaluation result. Only when the evaluation result meets the preset reliability standards will the pre-allocation strategy be approved for implementation.

[0024] Finally, the model update module ensures the system's adaptability. After the simulation-verified strategy is executed in a real cluster, this module collects feedback data such as actual task completion times and the actual resource consumption curves of each node. This data, along with earlier predictions and strategies, constitutes new training samples. The system employs online incremental learning algorithms, such as incremental learning based on random forests or online parameter update techniques for partial parameters of neural networks. Using these new samples, the internal parameters of the aforementioned prediction model and reinforcement learning strategy model are fine-tuned and optimized without service interruption, enabling the system to adapt to changes in task patterns and cluster environments. Through the coordinated operation of these four modules, the entire system forms a complete closed loop from multi-source data perception, intelligent prediction and self-correction, strategy simulation verification to continuous learning and optimization, achieving proactive judgment of computing power requirements and active pre-configuration of resources.

[0025] The demand forecasting and correction module includes a forecasting unit, a first detection unit, and a correction unit; The prediction unit is connected to the data acquisition and fusion module. It has a multi-head spatiotemporal attention neural network model built inside to receive the fused data and output a preliminary computing power demand prediction sequence. The multi-head spatiotemporal attention neural network model has 4 attention heads and 3 network layers. The GELU activation function is used, and the input dimension is 64-dimensional, consisting of historical computing power data, task type features, time features, and device status features. The output dimension is 1-dimensional, corresponding to the predicted computing power requirement. During model training, the mean squared error (MSE) is used as the loss function, and the Adam optimizer is selected. Gradient descent is used to minimize the deviation between the predicted and actual values, ensuring prediction accuracy.

[0026] The first detection unit is connected to the prediction unit and is used to perform real-time abnormal pattern recognition on the preliminary computing power demand prediction sequence based on an adaptive threshold algorithm. The abnormal patterns include numerical spikes, trend drifts, or periodic breaks. The adaptive threshold algorithm's initial value is set based on three standard deviations of historical computing power demand sequences, and the dynamic adjustment cycle remains consistent with the data acquisition cycle at 10 minutes. For numerical spike anomaly patterns, a rapid response mechanism is employed for threshold adjustment: when a sequence mutation is detected, the threshold is increased to 1.2 times the current sequence's maximum value within two data acquisition cycles. For trend drift anomaly patterns, the threshold is adjusted synchronously with the trend slope by linearly fitting recent sequence trends. For period break anomaly patterns, the threshold adjustment references the fluctuation range of similar periods, combining historical period data from the same period, ensuring the accuracy and timeliness of anomaly detection.

[0027] The correction unit is connected to the first detection unit and the prediction unit respectively. When the first detection unit identifies an anomaly, the correction unit is activated. It has a pre-set generative adversarial network model. The generator of the generative adversarial network model is used to reconstruct the identified abnormal data fragments to replace the original abnormal fragments and generate an intermediate correction sequence.

[0028] The generator of the Generative Adversarial Network (GAN) model employs a 3-layer convolutional structure, with batch normalization performed after each convolution. The dimensions of the fully connected layers are 256 and 128, respectively. The discriminator employs a 2-layer convolutional structure, with fully connected layers having dimensions of 256 and 1. The model outputs probability values. The training iterations are set to 3000. During training, cross-entropy is used as the GAN loss function, and the weight ratio between reconstruction loss and GAN loss is 1:1. Through adversarial training between the generator and discriminator, accurate correction of abnormal computing power demand sequences is achieved.

[0029] It should be further explained that the demand prediction and correction module is constructed and operated as follows: The prediction unit is implemented based on a specially designed deep learning architecture, which includes an embedding layer, multiple parallel attention heads, and a fully connected output layer. The embedding layer is responsible for mapping the multi-dimensional temporal feature vectors (including task history, system metrics, environmental labels, etc.) from the data acquisition and fusion module to a unified vector space.

[0030] Subsequently, multiple parallel attention heads work simultaneously. One attention head focuses on the temporal dependencies of the task submission sequence, another attention head focuses on the spatial correlations between system load metrics, and a third attention head is used to capture the impact patterns of external periodic events on the overall load.

[0031] The outputs of these attention heads are concatenated and normalized before being fed into a fully connected layer for regression calculation, ultimately outputting a preliminary prediction sequence representing the computing power demand values ​​at multiple consecutive future time points. The implementation of the first detection unit relies on a dynamically calculated anomaly detection mechanism. This unit maintains a reference window constructed based on the historical prediction error distribution of the prediction units and calculates in real time the deviation of each data point in the current preliminary prediction sequence from the local trend formed by its preceding and following adjacent points, as well as the deviation of that data point from the historical average of the same time period within the reference window.

[0032] The system presets an initial threshold, but this threshold is not fixed. Instead, it is dynamically adjusted based on the proportion of points recently marked as abnormal: when the proportion of abnormal points remains low, the threshold will be tightened appropriately to improve detection sensitivity; conversely, when abnormal alarms occur frequently, the threshold will be relaxed appropriately to prevent false alarms.

[0033] The core of the correction unit is a pre-trained generative adversarial network (GAN) model. During offline training, the generator of this GAN learns to reconstruct a smooth and reasonable demand curve from random noise and normal historical demand sequence fragments, while the discriminator learns to distinguish between real historical sequences and sequences generated by the generator.

[0034] In the online application phase, when the first detection unit identifies an anomalous segment (e.g., a sharp fluctuation in predicted values ​​across three consecutive time points), the generator is invoked. The generator takes the anomalous point itself and normal data from several time points before and after it as input, and outputs a new sequence segment of the same length that maintains contextual coherence, directly replacing the original anomalous data. This generative model-based reconstruction approach aims to repair the data at the data distribution level, rather than through simple interpolation or smoothing.

[0035] The demand forecasting and correction module also includes a second detection unit; The second detection unit is connected to the correction unit and is used to perform anomaly detection based on the adaptive threshold algorithm on the intermediate correction sequence again. If the second detection unit does not detect any abnormality, the intermediate correction sequence will be output as the final corrected computing power requirement sequence to the strategy generation and verification module. If the second detection unit detects an anomaly, it triggers the correction unit to reconstruct the new abnormal segment, or triggers a manual intervention process.

[0036] It should be further explained that the second detection unit is configured to perform an independent anomaly detection process on the intermediate corrected sequence output by the correction unit. This process is logically identical to that of the first detection unit, but the threshold parameters may be recalculated. This unit receives the data sequence reconstructed by the generative adversarial network model and uses the same composite judgment method based on local trend deviation and historical window reference values ​​to re-evaluate each data point in the sequence. Its adaptive threshold algorithm runs independently, dynamically adjusting based on the statistical characteristics (such as mean and variance) of the recent (e.g., within the past hour) intermediate corrected sequence and the frequency of correction operations being triggered, to ensure that the verification criteria for the corrected data are as rigorous as those for the original data.

[0037] If no data points are marked as abnormal during this detection, the system determines that the correction is effective and uses the current intermediate correction sequence as the final reliable corrected computing power requirement sequence, which is then directly passed to the subsequent strategy generation and verification module.

[0038] If the second detection unit still identifies new or incompletely corrected abnormal segments, the system will take further action according to the preset strategy: one path is to feed back the information of the newly discovered abnormal segment along with its context data to the correction unit again, and start a new round of reconstruction process for more specific segments; the other path is to determine that the abnormal pattern may exceed the processing capacity of the current automatic correction model, thereby triggering an alarm signal and recording relevant data snapshots, and transferring to the manual review and processing process by the system administrator.

[0039] This sequential design of "one correction, two verifications" forms a quality control loop with self-checking capabilities, aiming to ensure the integrity and reliability of the final output forecast data used for resource pre-allocation decisions to the greatest extent possible.

[0040] The strategy generation and verification module includes a strategy generation unit and a simulation verification unit; The strategy generation unit is connected to the demand prediction and correction module. It has a deep reinforcement learning model built inside. The deep reinforcement learning model takes the corrected computing power demand sequence as input and outputs a resource pre-allocation strategy that includes resource reservation, allocation ratio and scheduling time sequence. The state space of the deep reinforcement learning model is defined as an 8-dimensional vector, containing current system resource utilization, task queue length, historical average computing power requirement, task type distribution, node load status, resource idle rate, task priority ratio, and prediction error value. The action space is a discrete space, divided into 10 levels, corresponding to the quantization range of resource allocation ratios, with values ​​ranging from 0 to 1. The reward function is obtained through weighted summation, where the weight coefficient for task completion time is 0.4, the weight coefficient for load balancing is 0.3, and the weight coefficient for resource utilization is 0.3. The reward value guides the model to learn the optimal resource pre-allocation strategy.

[0041] The simulation verification unit is connected to the strategy generation unit. It has a lightweight system simulation environment built inside to simulate the execution of the resource pre-allocation strategy. Based on the load balancing index, task completion time index and resource utilization index generated during the simulation execution, the comprehensive evaluation score of the resource pre-allocation strategy is calculated.

[0042] The lightweight system simulation environment achieves its lightweight nature by simplifying non-core modules. At the hardware level, redundant monitoring and acquisition nodes are removed, retaining only core computing and communication nodes. At the software level, log storage and visualization modules are simplified, retaining only essential simulation data recording functions. The simulation time step is set to 1 second, consistent with the time granularity of the actual system. Node failure simulation employs a random triggering mechanism with a failure probability of 0.01, simulating a 50% drop in node computing power upon failure. Load surge simulation is based on historical peak data, randomly generating load demands 1.2 to 1.5 times the historical peak, with the surge lasting for 5 to 10 simulation time steps, thus simulating sudden load conditions in real-world scenarios.

[0043] In the calculation of the comprehensive evaluation score, the load balancing index is quantified using the coefficient of variation, which reflects the uniformity of load distribution by calculating the ratio of the standard deviation of the load of each node to the mean. The task completion timeliness index is quantified using the ratio of the actual completion time to the expected completion time; the closer the ratio is to 1, the better the timeliness performance. The resource utilization index is quantified using the ratio of the amount of resources used to the total amount of resources. The weight coefficients of each index are determined using the analytic hierarchy process (AHP). Users can adjust the weights of each index within the range of 0.1 to 0.5 according to the needs of the actual application scenario. The final comprehensive evaluation score is obtained by weighted summation of each quantified index and its corresponding weight.

[0044] It should be further explained that the policy generation and verification module operates as follows: The core of the policy generation unit is a deep reinforcement learning model, which is pre-trained to learn to make resource allocation decisions in a dynamic environment. The state space of this model is specifically designed as a vector containing the following elements: the predicted values ​​of the current and several future time windows in the verified and corrected computing power demand sequence provided by the demand prediction and correction module; the real-time system status obtained from the cluster manager, including the number of CPU cores, memory capacity, GPU memory, and network bandwidth currently available on each computing node; and the attributes of each task in the current task queue, such as estimated computing power, priority identifier, and dependencies.

[0045] The model's action space corresponds to a series of specific resource pre-allocation actions, such as "reserving X CPU cores and Y GB of memory on node A for high-priority tasks in the future Z time period", "adjusting the allocation ratio of a certain type of resource on node B to P%", or "suggesting that a batch of tasks be scheduled after time T".

[0046] The model's reward function is carefully designed, comprising several components: the reciprocal of task completion time (encouraging rapid processing), a resource utilization balancing factor (encouraging balanced use and avoiding hotspots), and a penalty term for violating resource constraints (such as over-allocation). The model learns and optimizes its policy network parameters through extensive trial-and-error interaction with the environment (initially a historical data simulation environment, later integrating with simulation verification units). Ultimately, it can output a specific and executable resource pre-allocation strategy for the corrected input demand sequence.

[0047] The simulation verification unit constructs an abstract yet comprehensive cluster behavior model. This model simulates the queue management logic of a real scheduler, the estimation of task execution time under different resource configurations, the data transmission latency between nodes limited by network bandwidth, and allows the injection of preset node failures or load surge events. Upon receiving the pre-allocated policy output by the policy generation unit, the simulation verification unit loads the policy, along with the current system snapshot and predicted requirements, into the sandbox environment and advances the simulation clock to simulate the policy's execution process within a virtual time period.

[0048] After the simulation, the unit extracts multiple quantitative metrics from the logs for calculation, including the average latency of all simulated tasks from submission to completion, the standard deviation of the overall cluster resource utilization (measuring balance), and the completion rate of high-priority tasks within the specified time limit. These metrics are aggregated using a predefined and configurable weighted formula to generate a comprehensive evaluation score ranging from 0 to 1, which is used to objectively measure the expected performance of the pre-allocation strategy under simulated conditions.

[0049] The simulation verification unit also forms a closed loop with the policy generation unit; If the overall evaluation score is higher than the preset qualification threshold, the resource pre-allocation strategy is allowed to be deployed and implemented in the real system. If the overall evaluation score is lower than the passing threshold, the evaluation results and corresponding scenario data will be fed back to the policy generation unit, triggering the deep reinforcement learning model to regenerate a new resource pre-allocation policy, or triggering manual policy adjustment.

[0050] It should be further explained that, in the closed-loop implementation, the simulation verification unit compares the calculated comprehensive evaluation score with a preset, configurable pass / fail threshold. This pass / fail threshold is preset by the system administrator based on the comprehensive requirements of the business for indicators such as scheduling reliability and resource utilization. If the score is greater than or equal to the threshold, it is determined that the current resource pre-allocation strategy performs as expected in the simulation environment, and the system generates an instruction to allow the strategy to be deployed to the real computing cluster for execution by subsequent modules. If the score is lower than the pass / fail threshold, it is determined that the strategy is insufficient in the current or simulated specific scenario.

[0051] At this point, the simulation verification unit not only outputs a failure signal, but also automatically packages a detailed evaluation report and scenario data package, which is sent to the policy generation unit as feedback. This data package contains the specific simulation scenario parameters that led to the low score, key event logs of the policy during simulation execution, and details of each sub-indicator. Upon receiving this feedback, the deep reinforcement learning model within the policy generation unit will initiate a policy optimization process according to preset rules. One mode is that the model uses the failed policy, its corresponding scenario, and the evaluation result as a new training sample, updates its internal network parameters, and immediately regenerates a new resource pre-allocation policy for the same prediction requirements based on the updated model. This new policy will then be submitted to the simulation verification unit for evaluation, forming an automated iterative optimization loop. Another mode is that when the continuously regenerated policies fail to reach the qualified threshold within a set number of iterations, or when the system detects that the evaluation failure stems from some extreme or novel scenarios that are difficult for the model to handle automatically, the system will automatically trigger an interrupt signal and generate a work order requiring manual review. The work order will submit the policy details, simulation results, and scenario data to the system administrator, who will then adjust the policy, redefine the rules, or correct the model training data, thereby intervening in the decision-making process.

[0052] The model update module includes a data collection unit and an incremental learning unit; The data collection unit is used to collect actual task execution time data, actual resource consumption data, and task execution status data after the resource pre-allocation strategy verified by the strategy generation and verification module is executed in the real system. The incremental learning unit is connected to the data collection unit and is used to update and optimize the parameters of the neural network model in the demand prediction and correction module and the deep reinforcement learning model in the policy generation and verification module without restarting the system by using actual task execution time data, real resource consumption data and task execution status data as new training samples.

[0053] The online incremental learning unit selects a partial parameter update method for neural networks. Specifically, the update scope includes the parameters of the fully connected and attention layers of a multi-head spatiotemporal attention neural network, as well as the value network parameters of a deep reinforcement learning model. The learning rate is set to 0.001, and the update cycle is synchronized with the system data acquisition cycle at 15 minutes. This unit employs an asynchronous update mechanism to achieve online optimization without restarting the system. By caching intermediate training results, model parameters are updated in batches. Each update only processes the gradients of parameters corresponding to newly added data, while limiting the size of parameters updated in a single batch to no more than 20% of the total parameters, effectively controlling memory usage and avoiding impact on system operation.

[0054] It should be further explained that the implementation process of the model update mechanism is as follows: After the data collection unit starts executing the resource pre-allocation strategy that has been verified by simulation in the real cluster, it starts continuous monitoring. This unit accurately collects the actual start time and end time of each executed task from the completion log of the cluster scheduler, and calculates the actual task execution time; at the same time, it collects the historical utilization curves of CPU, memory, GPU and network bandwidth corresponding to the strategy execution period from the resource monitoring agent of each node, as the actual resource consumption data; in addition, it also records the status information of whether the task was successfully completed, waited due to insufficient resources, or failed for other reasons.

[0055] The incremental learning unit employs an asynchronous processing pipeline. First, it aligns and correlates the collected real-world execution data with the corresponding data samples from the previous prediction phase, forming a complete empirical tuple containing "input features (original prediction basis), prediction output (preliminary prediction sequence), policy action (the pre-assigned policy executed), and actual results (actual time and cost)." Subsequently, the system utilizes an online learning algorithm based on mini-batch samples.

[0056] For the neural network model in the demand forecasting and correction module, incremental learning adopts a variant of the elastic weight solidification technique. Based on loading the original model parameters, it uses the input features in the new empirical tuple and the actual resource consumption results as supervision signals. Under the premise of limiting the drastic changes of important parameters, it performs gradient descent updates in a limited number of rounds to fine-tune the model weights so that its future predictions are more in line with the actual system feedback.

[0057] For the deep reinforcement learning model in the policy generation and verification module, the incremental learning unit stores the complete experience tuple as a new trajectory data in its experience replay buffer. It periodically samples from the buffer, mixing recent real interaction data with historical simulation data to jointly optimize and update the model's policy network and value network. This allows the agent to learn reward signals from the real environment and continuously adjust its policy. The entire update process is executed in a background thread without interrupting the system's main workflow.

[0058] Multi-source heterogeneous data includes task attribute data, system status data, external environment data, and real-time event stream data; Task attribute data includes task type, task priority, inter-task dependencies, and historical execution time records; System status data includes processor utilization data, memory usage data, network bandwidth usage data, and storage input / output data; External environment data includes user access pattern data and business cycle tag data; Real-time event stream data includes task submission queue data, system anomaly alarm log data, and resource release signal data.

[0059] It should be further explained that the specific content and collection method of the data are as follows: Task attribute data comes from the metadata of the job submission system or workflow engine. The task type is classified and identified according to its computational characteristics (such as CPU-intensive, memory-intensive, IO-intensive, or GPU-accelerated). The task priority is represented by a numerical level specified by the submitting user or automatically assigned according to business rules. The dependencies between tasks are described in the form of a directed acyclic graph, indicating the order of task execution. The historical execution time records are extracted from the log database of completed tasks, recording the actual running time of each task under different resource configurations.

[0060] System status data is obtained periodically by monitoring agents deployed on each computing node. Processor utilization data is the percentage of busy time for each logical core during the sampling period. Memory usage data is the ratio of used physical memory capacity to total capacity. Network bandwidth usage data is the number of bytes sent and received by the network interface per unit time. Storage input / output data is the number of read and write operations and the amount of data performed by the disk device per unit time.

[0061] User access pattern data in the external environment is analyzed by analyzing historical task submission time series to identify regular peak and trough periods on a daily and weekly basis; business cycle tag data is linked to an external calendar to mark weekdays, weekends, statutory holidays and specific business activity periods (such as e-commerce promotion seasons).

[0062] In the real-time event stream data, task submission queue data monitors the length of the pending queue and task attributes of the job scheduler in real time; system anomaly alarm log data collects key error and warning information from hardware monitoring, the operating system, and middleware; and resource release signal data is actively issued by the resource manager after a computing task is completed or a resource reservation timeout, indicating that a specific resource type and quantity on a specific node has been restored to an available state. These multi-source heterogeneous data, aligned using a unified timestamp, collectively constitute the comprehensive input for the system's prediction and decision-making.

[0063] A deep learning-based method for predicting computing power demand and pre-allocating resources, applied to a system, includes the following steps: S1: Collect multi-source heterogeneous data through the data acquisition and fusion module and perform preprocessing and feature fusion; S2: The demand forecasting and correction module generates a preliminary computing power demand forecasting sequence based on the fused data, performs at least one anomaly detection and reconstruction correction on the preliminary computing power demand forecasting sequence, and obtains the corrected computing power demand sequence. S3: The strategy generation and verification module generates a resource pre-allocation strategy based on the corrected computing power demand sequence, and verifies and evaluates the resource pre-allocation strategy in the simulation environment; S4: Deploy and execute the resource pre-allocation strategy that has passed the verification and evaluation in the real system; S5: Collects actual system operation data through the model update module and performs online incremental learning updates on the deep learning models involved in the system.

[0064] It should be further explained that the method follows a closed-loop sequence encompassing data acquisition, prediction correction, strategy verification, deployment execution, and continuous learning: First, the system initiates a data acquisition and fusion step, which actively obtains raw information from multiple pre-configured data interfaces, including job submission logs, cluster monitoring metrics, external time services, and system event streams. After format parsing, missing value imputation, and timestamp synchronization, the acquired raw data undergoes feature extraction and vectorization operations to fuse different types of business cycle labels, real-time resource metrics, and task metadata into a unified-dimensional time-series feature tensor, providing structured input for subsequent analysis.

[0065] Next, in the demand forecasting and correction step, the system calls the deployed multi-head spatiotemporal attention model to process the feature tensor and generate a preliminary computing power demand forecast sequence. This sequence then enters a preset anomaly handling process: the system first uses a dynamic threshold algorithm based on historical error distribution to scan and mark anomalies in the sequence, and then triggers a generative adversarial network model to reconstruct and replace the marked abnormal data segments to generate an intermediate sequence. This intermediate sequence needs to be verified again by the same anomaly detection logic to ensure the effectiveness of the correction, thereby completing the quality control closed loop of "at least one anomaly detection and reconstruction correction", and outputting the final corrected computing power demand sequence.

[0066] Subsequently, the policy generation and verification steps are initiated. The system inputs the corrected sequence along with the real-time resource status of the current cluster into the deep reinforcement learning model. Based on its internal policy network, the model outputs a draft pre-allocation policy containing specific resource reservation amounts, target nodes for allocation, and scheduling times. This draft policy is not executed immediately but is instead imported into a lightweight simulation environment for sandbox testing. The simulation environment simulates the scheduler behavior, network communication latency, and possible node failure scenarios of a real cluster and executes the draft policy.

[0067] After the simulation runs, the system automatically analyzes metrics such as task latency and resource utilization balance generated during the simulation, and calculates a quantitative comprehensive evaluation score. Only when this score meets a preset reliability threshold is the strategy allowed to proceed to the next stage. Then, the system submits the validated strategy to the cluster's resource manager or scheduler for actual deployment and execution.

[0068] Finally, in the model update step, the system continuously collects feedback data such as actual task completion time and resource consumption curves during the actual execution of the strategy, and associates these data with the corresponding prediction inputs and strategy actions to form training samples.

[0069] The system utilizes online learning algorithms to progressively adjust and optimize the parameters of the aforementioned prediction and strategy models using these new samples without affecting online services. This completes a full iteration from perception, decision-making, verification, execution to learning, and prepares for the next decision-making cycle.

[0070] A computer-readable storage medium stores a computer program that, when executed by a processor, implements a deep learning-based method for predicting computing power demand and pre-allocating resources. It should be further noted that the computer-readable storage medium can be a non-volatile memory such as a solid-state drive, optical disc, or flash memory chip. The computer program stored on this medium contains a series of machine instructions executable by a general-purpose or special-purpose processor. These instructions are organized into complete code that implements all the logical steps of the method. Specifically, the program code encapsulates software subroutines corresponding to the functional modules of data acquisition and fusion, demand prediction and correction, strategy generation and verification, and model updating.

[0071] The data acquisition and fusion subroutine defines the process and data format conversion rules for reading, parsing, and fusing multi-source heterogeneous data from a specified data source interface; the demand prediction and correction subroutine embeds the parameters of a trained multi-head spatiotemporal attention neural network model, the logic of an anomaly detection algorithm based on dynamic thresholds, and the generator weight file of a generative adversarial network model, and encodes the control flow of "prediction-detection-correction-re-verification"; the policy generation and verification subroutine encapsulates the policy network and value network parameters of a deep reinforcement learning model, as well as the construction and evaluation logic of a lightweight cluster simulation environment; and the model update subroutine implements the specific iterative steps and parameter update rules of the online incremental learning algorithm.

[0072] When the storage medium is connected to the computing device and read, the program stored on it is loaded into the device's memory, and the processor executes the above instructions one by one, thereby instantiating a closed-loop system that runs completely at the software level and has multi-level data processing and decision verification capabilities. This system can autonomously complete the entire process from data perception, demand prediction, anomaly correction, strategy simulation to continuous optimization according to the coded logic.

[0073] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement a deep learning-based method for predicting computing power demand and pre-allocating resources.

[0074] It should be further explained that the electronic device can be implemented as a physical server, a virtual machine instance, or a dedicated computing device. In this device, the memory is used to persistently store all computer program instructions for the implementation method and the supporting data required for system operation, including pre-trained neural network model parameter files, historical operation log databases, and system configuration files. After the device powers on, the processor loads the computer program from the memory into the main memory and begins executing instructions sequentially. These instructions drive the processor to perform a series of specific operations: First, the processor calls the data acquisition and fusion routine to periodically acquire multi-source data from external monitoring systems, job schedulers, and time services through the device's network interface or internal bus, and performs cleaning, alignment, and feature splicing in the memory to construct a time-series feature tensor in a unified format. Next, the processor loads the multi-head spatiotemporal attention model parameters into the computing core and performs forward propagation computation to generate a preliminary prediction sequence; subsequently, the processor runs an anomaly detection algorithm, compares the sequence data with dynamically maintained thresholds, and marks anomalies; if an anomaly is detected, the processor calls the loaded generative adversarial network generator to reconstruct and replace the anomaly fragment in the main memory, and performs secondary verification on the correction result.

[0075] During the policy generation phase, the processor loads a deep reinforcement learning model, inputs the validated demand sequence and the current system state vector into the model, and infers the specific resource pre-allocation action sequence through the policy network. Then, the processor starts a built-in simulation process to simulate the execution of the policy in an isolated memory space, and collects various performance metrics generated during the simulation for calculation and evaluation. After the evaluation is passed, the processor sends the policy instructions to the actual cluster resource manager via the network interface.

[0076] During policy execution, the processor simultaneously runs background data collection and model update threads, packaging the returned real-world data with corresponding historical decision data, and invoking incremental learning algorithms to fine-tune and update the parameters of the prediction and policy models residing in memory. Throughout this process, the memory provides instruction and data storage support, while the processor executes all computations, logical judgments, and control flows. Their collaborative work enables the electronic device to function as a complete physical carrier, achieving fully automated operation from multi-source data input, closed-loop prediction correction, policy simulation verification to autonomous continuous learning.

[0077] This system effectively addresses the dynamic and uncertain nature of computing power demands by constructing a prediction process that incorporates multi-source data fusion and features dual detection and correction mechanisms. This design ensures that prediction results maintain high reliability even when faced with data anomalies or sudden events, providing a stable and reliable data foundation for subsequent resource decisions and thus reducing the risk of scheduling errors caused by prediction biases from the outset.

[0078] Furthermore, by integrating strategy simulation verification and online incremental learning mechanisms, the system has transformed resource allocation from passive response to proactive pre-planning. This method can evaluate and optimize scheduling strategies before actual deployment, avoid potential bottlenecks in advance, and enable the system to continuously adjust itself based on actual operational feedback. This not only improves the efficiency and balance of cluster resource utilization but also enhances the system's adaptability to long-term task mode evolution and short-term emergency scenarios.

[0079] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0080] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based computing power demand prediction and resource pre-allocation system, characterized in that, It includes a data acquisition and fusion module, a demand forecasting and correction module, a strategy generation and verification module, and a model update module; The data acquisition and fusion module is used to acquire multi-source heterogeneous data and perform preprocessing and feature fusion on the data; The demand forecasting and correction module is connected to the data acquisition and fusion module. It is used to generate a preliminary computing power demand forecasting sequence based on the fused data, perform anomaly detection and correction on the preliminary computing power demand forecasting sequence, and output the corrected computing power demand sequence. The strategy generation and verification module is connected to the demand prediction and correction module, and is used to generate a resource pre-allocation strategy based on the corrected computing power demand sequence, and to verify and evaluate the resource pre-allocation strategy in a simulation environment. The model update module is connected to the data acquisition and fusion module, the demand prediction and correction module, and the strategy generation and verification module, respectively. It is used to collect actual system operation data and perform online incremental learning updates on the deep learning models involved in the system based on the actual operation data.

2. The computing power demand prediction and resource pre-allocation system based on deep learning according to claim 1, characterized in that: The demand forecasting and correction module includes a forecasting unit, a first detection unit, and a correction unit; The prediction unit is connected to the data acquisition and fusion module. It has a multi-head spatiotemporal attention neural network model built inside it to receive the fused data and output a preliminary computing power demand prediction sequence. The first detection unit is connected to the prediction unit and is used to perform real-time abnormal pattern recognition on the preliminary computing power demand prediction sequence based on an adaptive threshold algorithm. The abnormal patterns include numerical spikes, trend drifts, or periodic breaks. The correction unit is connected to the first detection unit and the prediction unit respectively. When the first detection unit identifies an anomaly, the correction unit is activated. It has a pre-set generative adversarial network model. The generator of the generative adversarial network model is used to reconstruct the identified abnormal data fragments to replace the original abnormal fragments and generate an intermediate correction sequence.

3. The deep learning-based computing power demand prediction and resource pre-allocation system according to claim 2, characterized in that: The demand forecasting and correction module also includes a second detection unit; The second detection unit is connected to the correction unit and is used to perform anomaly detection based on the adaptive threshold algorithm on the intermediate correction sequence again; If the second detection unit does not detect any abnormality, the intermediate correction sequence is output as the final corrected computing power requirement sequence to the strategy generation and verification module. If the second detection unit detects an anomaly, it triggers the correction unit to reconstruct the new abnormal segment, or triggers a manual intervention process.

4. The deep learning-based computing power demand prediction and resource pre-allocation system according to claim 3, characterized in that: The strategy generation and verification module includes a strategy generation unit and a simulation verification unit; The strategy generation unit is connected to the demand prediction and correction module. It has a deep reinforcement learning model built inside. The deep reinforcement learning model takes the corrected computing power demand sequence as input and outputs a resource pre-allocation strategy that includes resource reservation, allocation ratio and scheduling time sequence. The simulation verification unit is connected to the strategy generation unit. It has a lightweight system simulation environment built inside to simulate the execution of the resource pre-allocation strategy. Based on the load balancing index, task completion time index and resource utilization index generated during the simulation execution, the comprehensive evaluation score of the resource pre-allocation strategy is calculated.

5. The deep learning-based computing power demand prediction and resource pre-allocation system according to claim 4, characterized in that: The simulation verification unit also forms a closed loop with the strategy generation unit; If the overall evaluation score is higher than the preset qualification threshold, the resource pre-allocation strategy is allowed to be deployed and executed in the real system. If the comprehensive evaluation score is lower than the qualified threshold, the evaluation result and the corresponding scenario data are fed back to the policy generation unit, triggering the deep reinforcement learning model to regenerate a new resource pre-allocation policy, or triggering manual policy adjustment.

6. The computing power demand prediction and resource pre-allocation system based on deep learning according to claim 5, characterized in that: The model update module includes a data collection unit and an incremental learning unit; The data collection unit is used to collect actual task execution time data, actual resource consumption data, and task execution status data after the resource pre-allocation strategy verified by the strategy generation and verification module is executed in the real system. The incremental learning unit is connected to the data collection unit and is used to use the actual task execution time data, real resource consumption data, and task execution status data as new training samples to update and optimize the parameters of the neural network model in the demand prediction and correction module and the deep reinforcement learning model in the policy generation and verification module without restarting the system.

7. The computing power demand prediction and resource pre-allocation system based on deep learning according to claim 6, characterized in that: The multi-source heterogeneous data includes task attribute data, system status data, external environment data, and real-time event stream data; The task attribute data includes task type, task priority, inter-task dependencies, and historical execution time records; The system status data includes processor utilization data, memory usage data, network bandwidth usage data, and storage input / output data; The external environment data includes user access pattern data and business cycle tag data; The real-time event stream data includes task submission queue data, system anomaly alarm log data, and resource release signal data.

8. A method for predicting computing power demand and pre-allocating resources based on deep learning, characterized in that, Applied to the system of any one of claims 1 to 7, the method comprises the following steps: S1: Collect multi-source heterogeneous data through the data acquisition and fusion module and perform preprocessing and feature fusion; S2: The demand forecasting and correction module generates a preliminary computing power demand forecasting sequence based on the fused data, and performs at least one anomaly detection and reconstruction correction on the preliminary computing power demand forecasting sequence to obtain the corrected computing power demand sequence. S3: The strategy generation and verification module generates a resource pre-allocation strategy based on the modified computing power demand sequence, and verifies and evaluates the resource pre-allocation strategy in a simulation environment; S4: Deploy and execute the resource pre-allocation strategy that has passed the verification and evaluation in the real system; S5: Collects actual system operation data through the model update module and performs online incremental learning updates on the deep learning models involved in the system.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the deep learning-based computing power demand prediction and resource pre-allocation method as described in claim 8.

10. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the deep learning-based computing power demand prediction and resource pre-allocation method as described in claim 8.

Images