TTAO-LSTM less-data container cloud resource load prediction method
By using the TTAO-LSTM method, high-quality samples are generated by improving TimeGAN and TCN and multi-head attention mechanism, and the LSTM hyperparameters are optimized, which solves the problem of data scarcity in edge computing and achieves high-precision resource load prediction.
Patent Information
- Application Number
- CN202511790032.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-01
- Publication Date
- 2026-02-27
AI Technical Summary
In edge computing environments, due to limited hardware resources, container cloud devices cannot store high-resolution operational data for long periods, resulting in insufficient available training data, sparse, incomplete, and volatile load data, making it difficult for traditional models to accurately model load change trends.
The TTAO-LSTM method is adopted, which uses an improved TimeGAN generative model for data augmentation, combines a temporal convolutional network and a multi-head attention mechanism to generate high-quality samples, and optimizes the hyperparameters of the LSTM model through the TTAO algorithm to construct a TTAO-LSTM prediction model.
It significantly expands the scale of training data under limited data conditions, enhances the model's ability to extract temporal structure features, improves the accuracy and robustness of load prediction, and is suitable for lightweight deployment scenarios in edge computing and industrial internet.
Smart Images
Figure CN121579210A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of deep learning technology, and in particular to a TTAO-LSTM method for predicting resource load in low-data container clouds. Background Technology
[0002] In recent years, container cloud technology has gradually expanded to edge computing scenarios. Edge devices, constrained by both hardware costs and energy consumption, typically possess limited computing power and storage space. These resource limitations directly lead to the difficulty of edge nodes in storing high-resolution operational data for extended periods, resulting in a significant shortage of historical load data available for model training. This limited availability of training data samples poses a substantial challenge to cloud resource load prediction.
[0003] Load data in container cloud environments generally suffers from problems such as data sparsity, incompleteness, and drastic changes. On the one hand, the short deployment cycle and frequent instance start-up and shutdown of edge container services result in limited and highly volatile time-series data for continuous monitoring. On the other hand, the volatility of request traffic and the non-linear characteristics of business behavior cause load sequences to exhibit strong non-stationarity and multi-scale changes. These factors together make it difficult for traditional models to accurately model load change trends. Summary of the Invention
[0004] Therefore, it is necessary to provide a TTAO-LSTM method for predicting resource load in low-data container cloud, addressing the aforementioned technical issues.
[0005] The following technical solution is adopted in this specification: This manual provides a TTAO-LSTM method for predicting resource load in low-data container clouds, including: The original container cloud resource load data is normalized, and the normalized cloud resource load data is divided into multiple fixed-time series through a sliding window mechanism to construct a small sample load dataset. Based on the small sample workload dataset, data augmentation is performed on the small sample workload data using an improved TimeGAN generation model to generate augmented samples. The improved TimeGAN generation model replaces the RNN encoder by introducing a temporal convolutional network in the embedding module and forms the embedding domain by introducing a multi-head attention mechanism in the embedding-recovery module. Based on the enhanced samples, the LSTM model is trained to obtain a preliminary LSTM model; and the key hyperparameters of the preliminary LSTM model are searched and optimized using the TTAO algorithm to obtain the optimal hyperparameter combination, so as to construct the TTAO-LSTM prediction model. The cloud resource load is predicted using the TTAO-LSTM prediction model, and the predicted resource load value is obtained.
[0006] Furthermore, the normalization process for the original container cloud resource load data specifically includes: The resource load data is standardized using the max-min normalization method to eliminate the influence of different feature dimensions. The calculation formula is as follows: ; in, X represents the normalized load data; X represents the original load data. This indicates the minimum value of the load in the time series; This indicates the maximum value of the load in the time series.
[0007] Furthermore, the step of dividing the normalized cloud resource load data into multiple fixed-time series using a sliding window mechanism specifically includes: Assuming a window length of T and a step size of s, the normalized data is segmented to form a time series sample set; The event sequence sample set is represented as follows: ; in, N represents the i-th fixed-time sequence sample captured by the sliding window; N-T+1 represents the total number of sample sequences that can be divided according to the window length T when the total sequence length is N.
[0008] Furthermore, by introducing a temporal convolutional network instead of an RNN encoder in the embedding module, global trends in small sample load data are extracted; The temporal convolutional network expands the receptive field of the time series by dilating convolutions; it uses causal convolutions to ensure that the output depends only on the current and previous time steps, preventing time leakage. The dilated convolution is defined as: ; Where f(k) represents the convolution kernel weights; r is the dilation rate; and x(t) is the input time series. The causal convolution has the following constraints: ; in, Represents all possible time steps in a time series; Indicates at time step The input time series value at the specified location; t represents the time step corresponding to the current output.
[0009] Furthermore, the attention allocation process of the multi-head attention mechanism includes: The input vector X is linearly mapped into h distinct sets of query, key, and value vectors. ; For each group The self-attention output is calculated using the following formula: ; Where h is the number of attention heads; After concatenating the outputs of all attention points and performing a linear transformation, the calculation formula is as follows: in, These represent the query, key, and value vectors, respectively. It is a dimensional constant; To output linear transformation weights; By allocating attention to different time steps through a multi-head attention mechanism, the generated sequences can maintain temporal consistency and diversity, thereby enhancing the TimeGAN generative model's ability to allocate attention to different time steps.
[0010] Furthermore, the key hyperparameters of the LSTM model include the number of hidden layer neurons, learning rate, time step, and batch size.
[0011] Furthermore, the step of searching and optimizing the key hyperparameters of the LSTM model using the TTAO algorithm and constructing the TTAO-LSTM prediction model includes: TTAO algorithm initialization: Initialize the particle population, where each particle represents a set of hyperparameter combinations; Constructing a "topology graph": Each particle forms a triangular structure with its two topological neighbors to enhance local information exchange between populations; Define the objective function: use the RMSE of the LSTM model on the validation set as the fitness function; Triangular structure update strategy: The particle search direction is adjusted based on information sharing between triangular nodes through a local-global guided hybrid strategy; Output optimal parameters: After reaching the maximum number of iterations or fitness convergence, output the optimal combination of hyperparameters x∗ and train the final TTAO-LSTM prediction model; The calculation formula for the local-global guidance hybrid strategy is as follows: in, This represents the current hyperparameter vector of the i-th particle; , Its neighboring particles; It is the current global optimum; , This is the weighting factor.
[0012] This manual provides a TTAO-LSTM low-data container cloud resource load prediction system, including: Data preprocessing module: Normalizes the raw container cloud resource load data and divides the normalized cloud resource load data into multiple fixed-time series through a sliding window mechanism to construct a small sample load dataset; Time series generation module: Based on the small sample load dataset, the improved TimeGAN generation model is used to augment the small sample load data and generate augmented samples; wherein, the improved TimeGAN generation model introduces a temporal convolutional network to replace the RNN encoder in the embedding module; and introduces a multi-head attention mechanism in the embedding-recovery module to embed the network to form the embedding domain. TTAO-LSTM Model Training Module: Based on the augmented samples, the LSTM model is trained to obtain a preliminary LSTM model; and the key hyperparameters of the preliminary LSTM model are searched and optimized using the TTAO algorithm to obtain the optimal hyperparameter combination in order to construct the TTAO-LSTM prediction model. Load forecasting module: It uses the TTAO-LSTM forecasting model to forecast cloud resource load and obtain the resource load forecast value.
[0013] This specification provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described TTAO-LSTM method for predicting resource load in a low-data container cloud.
[0014] This specification provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the aforementioned TTAO-LSTM low-data container cloud resource load prediction method.
[0015] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects: This specification presents a TTAO-LSTM method for predicting resource load in container clouds with limited data. By introducing an improved TimeGAN time series generation model, along with a Temporal Convolutional Network (TCN) and a Multi-Head Attention mechanism, the time series generation model and the structure optimization model are organically integrated. This method can generate high-quality synthetic sequences with temporal consistency and structural diversity even with limited original data, significantly expanding the scale of training data and effectively alleviating the "sample scarcity" problem in container clouds at the edge. At the same time, it enhances the ability to extract temporal structural features, ensuring that the generated data retains the true fluctuation patterns while possessing good global trend perception capabilities, thus enhancing the model's modeling ability and prediction accuracy under limited data conditions.
[0016] Furthermore, the TTAO algorithm used in this invention achieves the cooperative evolution of population particles by constructing a local triangular connection structure, which enhances the local exploration ability in the early stage of the search and converges to the global optimal solution in the later stage of the search, effectively avoiding getting trapped in local optima. Attached Figure Description
[0017] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:
[0018] Figure 1 This document presents a flowchart illustrating a TTAO-LSTM method for predicting resource load in low-data container clouds. Figure 2 This document provides a flowchart illustrating an implementation of LSTM model optimization using the TTAO algorithm. Figure 3 This document provides a flowchart illustrating a TMA-TimeGAN sequence generation model based on a temporal convolutional network and a multi-head attention mechanism. Figure 4 This document provides a comparative analysis of experimental results for a TTAO-LSTM method for predicting resource load in low-data container clouds. Figure 5 The correlation t-sne analysis plot of the sequences generated by the TMA-TimeGAN model based on temporal convolutional networks and multi-head attention mechanism provided in this specification. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this application.
[0020] To address the problems of existing technologies, this invention proposes a TTAO-LSTM-based resource load prediction method for container clouds with limited data, based on time series generation. This method improves the accuracy and robustness of resource load prediction in scenarios with scarce data samples by combining an improved time series generation model, TimeGAN, with a triangular topology aggregation optimization-driven LSTM prediction model. Specifically, the TimeGAN model introduces a temporal convolutional network (TCN) and a multi-head attention mechanism to expand the receptive field of the time series and enhance the model's attention allocation to key time points, respectively, thereby generating higher-quality samples that are more realistic, structurally stable, and temporally consistent. The TTAO (triangular topology aggregation optimization) algorithm searches and optimizes key hyperparameters of the LSTM network, including the number of hidden units, learning rate, and sequence length, to construct the optimal load prediction model. Finally, it predicts future resource load trends and outputs high-precision results.
[0021] The method proposed in this invention aims to address the problem of insufficient resource prediction accuracy in edge container cloud platforms due to the scarcity of historical data samples and the strong non-stationarity of load sequences. This method organically integrates a time series generation model and a structural optimization model, enhancing the model's modeling ability and prediction accuracy under conditions of limited data.
[0022] It possesses strong versatility and scalability, and can be widely applied to small-sample resource load prediction tasks in edge computing, industrial internet, and lightweight deployment scenarios, thus having significant engineering value and application prospects.
[0023] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.
[0024] Figure 1 This is a flowchart illustrating a TTAO-LSTM method for predicting resource load in a low-data container cloud, as described in this specification. The method includes the following steps: S101: Data preprocessing module. To improve the comparability and stability of the model input data, the module performs max-min normalization on the original container cloud resource load data and divides the continuous time series into multiple fixed-length samples through a sliding window mechanism to form a standardized training set.
[0025] (1) The resource load data is standardized using the max-min normalization method to eliminate the influence of different feature dimensions: in, This represents the normalized load data; X represents the original load data (CPU utilization). This indicates the minimum value of the load in the time series; This indicates the maximum value of the load in the time series.
[0026] (2) Perform sliding window processing. Assuming the window length is T and the step size is s, segment the normalized data to form a time series sample set: in, This indicates the first segment captured by the sliding window. i A fixed-time series sample; N-T+1 This represents the total number of sample sequences that can be divided based on the window length T, given a total sequence length of N.
[0027] S102: Time series generation module, which adopts an improved TimeGAN generation model to enhance time series samples based on small sample size; and improves the temporal consistency and structural authenticity of the generated data by introducing a TCN network and a multi-head attention mechanism.
[0028] (1) Embedded module introduced into TCN network The TMA-TimeGAN generative model is used to augment small sample data. The TMA-TimeGAN model introduces a temporal convolutional network (TCN) and a multi-head attention mechanism, which are used to expand the receptive field of the time series and enhance the model's attention allocation to key time points, respectively.
[0029] Using a temporal convolutional network instead of a traditional RNN encoder, its dilated convolution and causal convolution forms are as follows: Dilated convolution is defined as in, For convolution kernel weights, For expansion rate, For input time series.
[0030] Causal convolution ensures that the output depends only on the current and previous time steps, preventing information leakage.
[0031] Its basic constraints are: .
[0032] in, This represents all possible time steps in a time series; Indicates at time step The input time series value at the location.
[0033] TCN has a larger receptive field, which helps to extract global trends in time series.
[0034] (2) The embedded-recovery module introduces Multi-Head Attention The preprocessed data is fed into the sequence generator. By embedding a temporal convolutional network with a multi-head attention mechanism into the generator, a generated embedding space with more similar features to the original data embedding space is generated. The original data embedding space and the generator-generated embedding space are fed into the discriminator for discrimination. Finally, the trained embedding space is fed into the recovery network containing the temporal convolutional network with a multi-head attention mechanism to generate enhanced samples.
[0035] To enhance the model's ability to allocate attention at different time steps, a multi-head attention mechanism is introduced; The input vector X is linearly mapped into h distinct sets of query, key, and value vectors. Where h is the number of attention heads, for each group Calculate the self-attention output: Concatenate the outputs of all attention points and then perform a linear transformation: in, These represent the query, key, and value vectors, respectively. It is a dimensional constant; The weights are for the linear transformation of the output.
[0036] Finally, sample data with similar temporal structure and statistical characteristics to the original load data are generated to expand the training set and improve the model's generalization ability.
[0037] S103: TTAO-LSTM model training module, which uses the TTAO algorithm to search and optimize the key hyperparameters of the LSTM model and build the optimal prediction model.
[0038] The key hyperparameters of the LSTM model include: the number of neurons in the hidden layer. hidden Learning rate φ, time step φ, batch size φ; (1) TTAO algorithm initialization: Initialize the particle population (each particle represents a set of hyperparameter combinations); Constructing a "topological graph": Each particle forms a triangular structure with its two topological neighbors to enhance local information exchange among populations; Define the objective function: use the RMSE of the LSTM model on the validation set as the fitness function.
[0039] (2) Triangular structure update strategy: By adjusting the particle search direction through information sharing between triangular nodes, a hybrid local-global guidance strategy is adopted. in, This represents the current hyperparameter vector of the i-th particle; , Its neighboring particles; It is the current global optimum; , This is the weighting factor.
[0040] (3) Output optimal parameters: After reaching the maximum number of iterations or fitness convergence, output the optimal hyperparameter combination x. ∗ The final TTAO-LSTM model was trained.
[0041] S104: Load prediction and evaluation module. Input test samples into the trained TTAO-LSTM model, output predicted values, and evaluate the prediction accuracy through RMSE and MAE metrics to complete the entire load prediction process.
[0042] The trained TTAO-LSTM model is used to predict the test data, outputting predicted resource load values. The accuracy of the prediction results was evaluated using the following metrics:
[0043] (1) Root Mean Square Error (RMSE): (2) Mean Absolute Error (MAE): in, For the true value, For predicted values, This represents the number of test samples.
[0044] like Figure 2 As shown, the process of optimizing an LSTM model using the TTAO algorithm according to this invention consists of the following steps: Step 1: Initialize the population, with each individual representing a set of LSTM hyperparameters; Step 2: Define the fitness function and use the RMSE on the training set as the evaluation criterion; Step 3: Calculate the fitness value to obtain the individual's relationship with the current global optimum; Step 4: Construct a triangular topology by selecting two neighbors for each individual to form a triangular relationship; Step 5: Update parameters based on topological relationships, and adjust particle hyperparameters by reference point weighting. Step 6: Determine the termination condition. If the maximum algebra or convergence condition is met, output the current best individual to form the final TTAO-LSTM model. like Figure 3 As shown, the generation of enhanced samples for the TMA-TimeGAN model based on temporal convolutional networks and multi-head attention mechanisms in this invention consists of the following steps: First, the original payload data is processed by a convolutional embedding network to extract the time series embedding representation, thus constructing the original data embedding space; Then, the generator takes random noise as input and passes it through a multi-head attention mechanism and a temporal convolutional network in sequence to generate an embedding representation with temporal consistency and key point attention capabilities, thus obtaining the generated embedding space.
[0045] Next, the generated embedding space and the original embedding space are input together into the discriminator for adversarial training, so that the generator learns the distribution characteristics of the original data and improves the authenticity of the generated samples.
[0046] After adversarial training, the generated embedding representations are fed into a recovery network that incorporates a multi-head attention mechanism and a temporal convolutional network. This ultimately generates augmented samples with similar statistical properties and temporal structure to the real payload data.
[0047] like Figure 4 The figure shown is a comparative analysis of experimental results of the TTAO-LSTM method for predicting resource load in container clouds based on time series generation, according to the present invention.
[0048] The solid line represents the actual CPU utilization load curve, while the dashed line represents the CPU utilization load value predicted by the LSTM model optimized with TMA-TimeGAN data augmentation and TTAO. The overall trend shows that the predicted curve closely follows the actual data trajectory, maintaining a high degree of fit in most time steps, and effectively capturing the dynamic characteristics of the actual sequence, especially in key regions such as peak abrupt changes and frequent fluctuations.
[0049] The dashed predicted curve and the solid actual curve show a consistent overall trend, indicating that the model has a strong ability to capture global trends. Even in regions with drastic load fluctuations (such as steps 100, 200, and 300), the model can still respond quickly and accurately reproduce the changes in the original data, demonstrating its superior robustness in non-stationary, highly dynamic scenarios. The samples generated by TMA-TimeGAN enrich the training dataset, effectively alleviating the bottleneck of small-sample learning on edge devices, thereby improving the model's generalization ability and stability. The TTAO algorithm effectively optimizes the structural hyperparameters of the LSTM network (such as the number of hidden layers, learning rate, and sequence length), further improving prediction accuracy and convergence speed.
[0050] Figure 5 The correlation t-sne analysis diagram of the sequences generated by the TMA-TimeGAN model based on temporal convolutional networks and multi-head attention mechanism in this invention. In the figure, black dots represent the distribution of the original data in the embedding space, and gray dots represent the embedding results of the generated data. The t-SNE (t-distributed random neighborhood embedding) algorithm is used to perform dimensionality reduction mapping on the high-dimensional temporal embedding features, thereby intuitively presenting the data distribution characteristics and similarities in a two-dimensional space.
[0051] As can be observed from the figure, the generated data and the original data under grayscale markings exhibit significant mixed clustering and morphological overlap in most areas, indicating that the TMA-TimeGAN generated sequences demonstrate good consistency and reliability in capturing the statistical structure and temporal characteristics of the original data. Grayscale contrast further highlights the commonalities in data clustering and spatial structure, verifying the indistinguishability of the generated samples and the original samples in the embedding space, which is a sign of high-quality generation.
[0052] In summary, this invention provides a TTAO-LSTM-based method for predicting resource load in container clouds with limited data, based on time series generation. This method improves the accuracy and robustness of resource load prediction in scenarios with scarce data samples by combining the improved time series generation model TimeGAN with a triangular topology aggregation optimization-driven LSTM prediction model. Specifically, the TimeGAN model introduces a temporal convolutional network (TCN) and a multi-head attention mechanism to expand the receptive field of the time series and enhance the model's attention allocation to key time points, respectively, thereby generating higher-quality samples that are more realistic, structurally stable, and temporally consistent. The TTAO (triangular topology aggregation optimization) algorithm searches and optimizes key hyperparameters of the LSTM network, including the number of hidden units, learning rate, and sequence length, to construct the optimal load prediction model. Finally, it predicts future resource load trends and outputs high-precision results. The proposed method has strong versatility and scalability, and can be widely applied to resource load prediction tasks with limited data in edge computing, industrial internet, and lightweight deployment scenarios, demonstrating significant engineering value and application prospects.
[0053] The present invention has the following beneficial effects and advantages: (1) Compared with traditional load prediction methods based on large sample training, this invention has the ability to efficiently model in small sample scenarios. This invention introduces an improved TimeGAN time series generation model, which can generate high-quality synthetic sequences with temporal consistency and structural diversity when the amount of original data is limited, significantly expanding the scale of training data and effectively alleviating the "sample scarcity" problem of container clouds in edge environments. This generation model enhances the ability to extract temporal structural features by introducing a temporal convolutional network (TCN) and a multi-head attention mechanism, ensuring that the generated data retains the real fluctuation pattern while having a good global trend perception ability, providing richer input features for subsequent prediction models.
[0054] (2) Compared with traditional fixed-parameter LSTM models, this invention introduces the TTAO optimization algorithm to achieve automatic optimization of LSTM hyperparameters, thereby improving the prediction accuracy and stability of the model. The Triangular Topological Aggregation Optimization Algorithm (TTAO) used in this invention has strong search capabilities and a topological information fusion mechanism. TTAO achieves the co-evolution of population particles by constructing a local triangular connection structure, enhancing local exploration capabilities in the early stage of the search and converging to the global optimum in the later stage of the search, effectively avoiding getting trapped in local optima. Compared with conventional optimization methods such as Genetic Algorithm (GA) and Particle Swarm Optimization (PSO), TTAO can still efficiently search for the optimal hyperparameter combination when the parameter space dimension is high and the data distribution is complex, thereby significantly improving the adaptability and generalization ability of the LSTM model in different data environments.
[0055] (3) This invention has good adaptability and flexibility, and is suitable for edge computing and resource-constrained scenarios. Since this method can still achieve high-precision prediction under the condition of a small number of samples, and the structure of each module is lightweight and the computational load of the algorithm is moderate, it can be widely used in real-time load prediction tasks of container cloud in edge devices, IoT gateways and embedded deployment environments, providing a reliable basis for container scheduling, resource elastic configuration and so on.
[0056] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
Claims
1. A TTAO-LSTM method for predicting resource load in low-data container clouds, characterized in that, include: The original container cloud resource load data is normalized, and the normalized cloud resource load data is divided into multiple fixed-time series through a sliding window mechanism to construct a small sample load dataset. Based on the small sample workload dataset, data augmentation is performed on the small sample workload data using an improved TimeGAN generation model to generate augmented samples. The improved TimeGAN generation model replaces the RNN encoder by introducing a temporal convolutional network in the embedding module and forms the embedding domain by introducing a multi-head attention mechanism in the embedding-recovery module. Based on the enhanced samples, the LSTM model is trained to obtain a preliminary LSTM model; and the key hyperparameters of the preliminary LSTM model are searched and optimized using the TTAO algorithm to obtain the optimal hyperparameter combination, so as to construct the TTAO-LSTM prediction model. The cloud resource load is predicted using the TTAO-LSTM prediction model, and the predicted resource load value is obtained.
2. The TTAO-LSTM method for predicting resource load in a low-data container cloud as described in claim 1, characterized in that, The normalization process for the original container cloud resource load data specifically includes: The resource load data is standardized using the max-min normalization method to eliminate the influence of different feature dimensions. The calculation formula is as follows: ； in, X represents the normalized load data; X represents the original load data. This indicates the minimum value of the load in the time series; This indicates the maximum value of the load in the time series.
3. The TTAO-LSTM method for predicting resource load in low-data container clouds as described in claim 1, characterized in that, The process of dividing the normalized cloud resource load data into multiple fixed-time series using a sliding window mechanism specifically includes: Assuming a window length of T and a step size of s, the normalized data is segmented to form a time series sample set; The event sequence sample set is represented as follows: ； in, N represents the i-th fixed-time sequence sample captured by the sliding window; N-T+1 represents the total number of sample sequences that can be divided according to the window length T when the total sequence length is N.
4. The TTAO-LSTM method for predicting resource load in a low-data container cloud as described in claim 1, characterized in that, By introducing a temporal convolutional network instead of an RNN encoder in the embedding module, global trends in small sample load data can be extracted. The temporal convolutional network expands the receptive field of the time series by dilating convolutions; it uses causal convolutions to ensure that the output depends only on the current and previous time steps, preventing time leakage. The dilated convolution is defined as: ； Where f(k) represents the convolution kernel weights; r is the dilation rate; and x(t) is the input time series. The causal convolution has the following constraints: ； in, Represents all possible time steps in a time series; Indicates at time step The input time series value at the specified location; t represents the time step corresponding to the current output.
5. The TTAO-LSTM method for predicting resource load in a low-data container cloud as described in claim 1, characterized in that, The attention allocation process in multi-head attention mechanisms includes: The input vector X is linearly mapped into h distinct sets of query, key, and value vectors. ; For each group The self-attention output is calculated using the following formula: ； Where h is the number of attention heads; After concatenating the outputs of all attention points and performing a linear transformation, the calculation formula is as follows: in, These represent the query, key, and value vectors, respectively. It is a dimensional constant; To output linear transformation weights; By allocating attention to different time steps through a multi-head attention mechanism, the generated sequences can maintain temporal consistency and diversity, thereby enhancing the TimeGAN generative model's ability to allocate attention to different time steps.
6. The TTAO-LSTM method for predicting resource load in a low-data container cloud as described in claim 1, characterized in that, The key hyperparameters of the LSTM model include the number of hidden layer neurons, learning rate, time step, and batch size.
7. The TTAO-LSTM method for predicting resource load in a low-data container cloud as described in claim 1, characterized in that, The process of searching and optimizing key hyperparameters of the LSTM model using the TTAO algorithm and constructing a TTAO-LSTM prediction model includes: TTAO algorithm initialization: Initialize the particle population, where each particle represents a set of hyperparameter combinations; Constructing a "topology graph": Each particle forms a triangular structure with its two topological neighbors to enhance local information exchange between populations; Define the objective function: use the RMSE of the LSTM model on the validation set as the fitness function; Triangular structure update strategy: The particle search direction is adjusted based on information sharing between triangular nodes through a local-global guided hybrid strategy; Output optimal parameters: After reaching the maximum number of iterations or fitness convergence, output the optimal combination of hyperparameters x∗ and train the final TTAO-LSTM prediction model; The calculation formula for the local-global guidance hybrid strategy is as follows: in, This represents the current hyperparameter vector of the i-th particle; , Its neighboring particles; It is the current global optimum; , This is the weighting factor.
8. A TTAO-LSTM low-data container cloud resource load prediction system, characterized in that, include: Data preprocessing module: Normalizes the raw container cloud resource load data and divides the normalized cloud resource load data into multiple fixed-time series through a sliding window mechanism to construct a small sample load dataset; Time series generation module: Based on the small sample load dataset, the small sample load data is augmented using an improved TimeGAN generation model to generate augmented samples; wherein, the improved TimeGAN generation model replaces the RNN encoder by introducing a temporal convolutional network in the embedding module. An embedding domain is formed by introducing a multi-head attention mechanism into the embedding-recovery module into the network. TTAO-LSTM Model Training Module: Based on the augmented samples, the LSTM model is trained to obtain a preliminary LSTM model; and the key hyperparameters of the preliminary LSTM model are searched and optimized using the TTAO algorithm to obtain the optimal hyperparameter combination in order to construct the TTAO-LSTM prediction model. Load forecasting module: It uses the TTAO-LSTM forecasting model to forecast cloud resource load and obtain the resource load forecast value.
9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1 to 7.
10. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any one of claims 1 to 7.