Resource demand analysis method based on multi-scale deep learning prediction
By using a multi-scale deep learning prediction method combined with the Cross-Attention and BiLSTM models to analyze resource demand, the real-time and efficiency issues of resource scheduling in existing technologies are solved, efficient resource demand prediction and intelligent scheduling are achieved, and the system's response speed and resource utilization are improved.
Patent Information
- Application Number
- CN202510823877.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-19
- Publication Date
- 2025-09-16
AI Technical Summary
Existing deep learning algorithms have problems in resource scheduling, such as long training time, high computational consumption, overfitting, and gradient vanishing, making it difficult to meet real-time and high efficiency requirements. In particular, resource scheduling strategies are not responsive enough in dynamic environments. Traditional LSTM models have high computational complexity when processing large-scale data, leading to performance bottlenecks.
A multi-scale deep learning prediction method is adopted, including extracting network logs based on Python program and saving them as CSV files, using the Cross-Attention mechanism to generate state representation vectors, combining the BiLSTM model for time series prediction, and making resource scheduling decisions through the PPO algorithm to achieve intelligent resource allocation.
It improves the accuracy of resource demand forecasting and dynamic adjustment capabilities, enhances the model's ability to understand complex patterns, supports real-time monitoring and resource optimization, reduces operation and maintenance costs, and improves the robustness and efficiency of the system.
Smart Images
Figure CN120653402A_ABST
Abstract
Claims
1. A resource demand analysis method based on multi-scale deep learning prediction, characterized in that: include: Step 1: Extract computer network logs based on Python program and save them as CSV files; Step 2: Using the Cross-Attention mechanism, we fuse computer network logs to generate a unified state representation vector. Step 3: Time series prediction based on the BiLSTM model; Step 4: Make resource scheduling decisions based on the state representation vector and prediction results, and allocate computing resources reasonably.
2. The method according to claim 1, characterized in that In step 1, the computer network log is extracted based on the Python program and saved as a CSV file. Specifically: The system is monitored based on a Python program. The system resource usage, including CPU usage, memory usage, and network statistics, is regularly collected. The latest monitoring data is printed on the console and finally saved in a CSV file.
3. The method according to claim 2, characterized in that In step 2, the Cross-Attention mechanism is used to fuse computer network logs to generate a unified state representation vector, specifically: Read CSV files and normalize CPU usage, memory usage, and network statistics; A multimodal fusion model is constructed to perform multimodal fusion on the standardized CPU usage, memory usage, and network statistics, and output a state representation vector.
4. The method according to claim 3, characterized in that In step 3, time series prediction is performed based on the BiLSTM model, specifically: Read CPU usage, memory usage, and network statistics, perform window sliding processing on them, and generate a format suitable for time series model training; Build a BiLSTM model and train it based on the data after window sliding processing; Based on the trained BiLSTM model, use the predict_point_by_point() function to implement point-by-point prediction. The input is the latest time window data, and the model outputs resource trend forecasts for the next prediction_length steps. After each prediction, the window is slid forward one position to implement a rolling forecast for the entire future time period. The prediction results are saved in CSV format for use by the reinforcement learning module for intelligent scheduling.
5. The method according to claim 4, characterized in that In step 4, resource scheduling decisions are made based on the state representation vector and prediction results to reasonably allocate computing resources. Specifically: The resource scheduling problem is modeled based on the PPO algorithm to obtain a resource scheduling decision model; Resource scheduling decisions are made through a resource scheduling decision model based on the state representation vector and prediction results.
6. The method according to claim 5, characterized in that The resource scheduling problem is modeled based on the PPO algorithm, specifically: Define the state space, action space, and reward function. For the state space, construct a feature vector by combining real-time monitoring data and prediction model output. Actions correspond to different resource allocation plans. By collecting empirical data, i.e., state, action, and reward sequences, the strategy is continuously updated. PPO introduces a clipping objective function to ensure that the new and old strategies do not differ too much during the update process, thereby balancing exploration and exploitation. Regularly evaluate the performance of the model to check whether it has achieved the expected goals, and adjust the model structure or training parameters based on the evaluation results.