System resource demand planning method and system based on artificial intelligence
By building a heterogeneous graph of resource attributes and resource demand prediction network, combining graph embedding models and iterative community optimization, the accuracy and multi-dimensional consideration of resource demand planning in the existing technology are solved, and accurate prediction and optimization of system resource demand are achieved. , improve resource utilization efficiency and service quality.
Patent Information
- Application Number
- CN202410994419.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-24
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-07-24
AI Technical Summary
The existing technology is difficult to accurately predict and optimize the complex system resource requirements of data centers and cloud computing platforms, resulting in low resource utilization and degradation of service quality, and lack of comprehensive consideration and collaborative optimization of multi-dimensional resources.
By constructing a heterogeneous graph of resource attributes, using multiple rounds of random walks to generate resource node sequences, training graph embedding models to learn node neighborhood relationships, generating candidate feature matrix and optimal feature subsets, and combining resource demand prediction network capture time-step dependencies to predict network capture Relationship, generate mixed Gaussian distribution parameters, randomly generate planning schemes, and iterative community optimization to finally determine the optimal resource planning scheme.
It realizes accurate prediction and dynamic planning of system resource requirements, improves resource utilization efficiency, ensures service quality, reduces operation and maintenance costs, makes full use of the advantages and complementarity of various models, and improves prediction and optimization performance.
Smart Images

Figure CN118521139B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an artificial intelligence-based system resource demand planning method and system. Background Art
[0002] With the rapid development of information technology and the continuous expansion of its application scale, data centers and cloud computing platforms are facing increasingly complex and dynamically changing system resource requirements. How to timely and accurately predict future resource requirements and formulate reasonable resource planning plans based on changes in business load and resource utilization has become a key issue in improving resource utilization efficiency and ensuring service quality.
[0003] Traditional resource demand planning methods mainly rely on manual experience and simple statistical models. When faced with complex and nonlinear resource demand patterns, it is often difficult to capture the key features and trends, resulting in low prediction accuracy and difficulty in meeting actual needs. At the same time, they lack the ability to adjust dynamically changing business loads and resource status in real time, resulting in low resource utilization and reduced service quality. The resource demand planning method based on artificial intelligence also has some shortcomings. It lacks comprehensive consideration and collaborative optimization of multi-dimensional resources, and it is difficult to fully utilize the advantages and complementarity of different models, which limits the performance improvement space for prediction and optimization.
[0004] To sum up, there is a need for a system resource demand planning method based on artificial intelligence that can comprehensively consider the demand characteristics of multi-dimensional resources, utilize the collaborative optimization of multiple machine learning models, and realize accurate prediction and dynamic planning of resource demand, thereby improving resource utilization efficiency, ensuring service quality, and reducing operation and maintenance costs. The present invention can solve the problems in the prior art. Summary of the invention
[0005] The embodiments of the present invention provide a system resource demand planning method and system based on artificial intelligence, which can solve the problems in the prior art.
[0006] According to a first aspect of the embodiments of the present invention,
[0007] Provided is an artificial intelligence-based system resource requirement planning method, including:
[0008] Acquire system resource data, construct a resource attribute heterogeneous graph, generate a resource node sequence set through multiple rounds of random walks based on the resource attribute heterogeneous graph, learn the node neighborhood relationship through a training graph embedding model, construct a candidate feature matrix, and generate an optimal feature subset based on the comprehensive feature score of each candidate feature;
[0009] The system resource time series data to be predicted is input into the resource demand prediction network, and the mixed Gaussian distribution parameters are generated by capturing the dependency relationship of each time step, and the conditional probability distribution is determined to obtain the resource demand prediction result;
[0010] According to the resource demand prediction results and the system resources determined based on the optimal feature subset, a plurality of planning schemes are randomly generated, an iterative community is constructed, and the iterative community is input into the resource allocation algorithm to generate the optimal solution and determine the optimal planning scheme.
[0011] In an optional embodiment,
[0012] Obtain system resource data, build a resource attribute heterogeneous graph, generate a resource node sequence set through multiple rounds of random walks based on the resource attribute heterogeneous graph, learn the node neighborhood relationship through the training graph embedding model, build a candidate feature matrix, and generate the optimal feature subset based on the comprehensive feature score of each candidate feature, including:
[0013] Obtain system resource data, determine resource entities and resource attributes, use resource entities and resource attributes as resource nodes, and use association weights between resource nodes as connecting edges to construct a resource heterogeneous graph;
[0014] Randomly select a central node from the resource heterogeneous graph, determine the context node through random walk, and generate a resource node sequence;
[0015] After multiple rounds of random walks, multiple resource node sequences are generated to form a resource node sequence set;
[0016] Taking the resource node sequence set as data samples, the central node as input, and the context node as output, based on the pre-built graph embedding model, learning the node neighborhood relationship, generating a low-dimensional dense vector representation corresponding to each node, and combining them into a candidate feature matrix;
[0017] Based on each column of feature vectors in the candidate feature matrix, the feature comprehensive scores are calculated and sorted, and according to the preset optimal number of features, the corresponding number of candidate features are selected to generate the optimal feature subset.
[0018] In an optional embodiment,
[0019] From the resource heterogeneous graph, randomly selecting a central node, determining a context node through random walk, and generating a resource node sequence includes:
[0020] From the resource heterogeneous graph, randomly select the resource node corresponding to the resource entity as the starting node, take the starting node as the current node, and start random walking:
[0021] Based on the current node, calculate the node degree, determine the return factor, and calculate the neighborhood clustering coefficient to determine the exploration factor. Combine the weights of all the edges connected to the current node to determine the transition probability distribution corresponding to the neighboring node. Based on the transition probability distribution, set a random roulette wheel to determine the next node and add it to the node sequence.
[0022] The next node is set as the current node, and the random walk is repeated until the length of the node sequence reaches a preset sequence length, thereby generating a resource node sequence.
[0023] In an optional embodiment,
[0024] Based on the current node, the node degree is calculated, the return factor is determined, and the clustering coefficient of the neighborhood is calculated, the exploration factor is determined, and the weights of all the connecting edges connected to the current node are combined to determine the transition probability distribution corresponding to the neighboring node, including:
[0025] The transition probability from the current node to the neighboring node is as follows:
[0026] ;
[0027] in, u Representation Node u , v Represents another node v , P ( u | v ) indicates a slave node v To Node u The transition probability, β represents the balance factor, deg( v ) represents a node v The node degree of N ( v ) represents a node v The set of neighboring nodes of x Represents a node in the neighborhood node set x , C u Representation Node u The clustering coefficient of the neighborhood, C x Representation Node x The clustering coefficient of the neighborhood, ω vu Representation Node v and nodes u The weight of the edge connecting them, y Represents a node in the neighborhood node set y , C y Representation Node yThe clustering coefficient of the neighborhood, ω vx Representation Node v and nodes x The weight of the connecting edge.
[0028] In an optional embodiment,
[0029] The system resource time series data to be predicted is input into the trained optimal resource demand prediction network. By capturing the dependency of each time step, the mixed Gaussian distribution parameters are generated, and the conditional probability distribution is determined. The resource demand prediction results include:
[0030] Acquire time series data of system resources, preprocess the time series data, divide the time series data into multiple groups based on a preset sliding window size, determine an input sequence and an output sequence, and generate a training sample set;
[0031] Constructing a resource demand prediction network, including an input layer, at least one recursive hidden layer and an output layer, wherein the input layer receives the input sequence; the recursive hidden layer is constructed based on a long short-term memory network, and controls the information transmission of the input sequence through a gating mechanism to capture long-term dependencies; the output layer includes a plurality of fully connected layer structures, and generates mixed Gaussian distribution parameters through direct output, exponential function transformation output and activation function normalized output respectively;
[0032] Inputting the training sample set into the resource demand prediction network, training based on the gradient descent method, iteratively adjusting the model parameters by maximizing the loss function until a preset number of iterations is reached to generate an optimal resource demand prediction network;
[0033] The time series data to be predicted is input into the optimal resource demand prediction network to obtain mixed Gaussian distribution parameters, determine the conditional probability distribution, and determine the resource demand prediction result by calculating the expectation of the conditional probability distribution.
[0034] In an optional embodiment,
[0035] By maximizing the loss function and iteratively adjusting the model parameters until the preset number of iterations is reached, the optimal resource demand prediction network is generated, including:
[0036] The loss function has the following formula:
[0037] ;
[0038] in, L represents the loss function, i represents the Gaussian component ordinal number, M represents the total number of mixed distributions, πi Indicates i The mixing coefficients of Gaussian components, p represents the variables in the input sequence, q represents the variables in the output sequence, μ i Indicates i The mean of the Gaussian components, σ i Indicates i The standard deviation of the Gaussian components.
[0039] In an optional embodiment,
[0040] The iterative process of the resource allocation algorithm includes:
[0041] Taking the planning scheme as an individual, initialize multiple populations and construct an iterative community;
[0042] Based on the preset fitness function, the fitness value of each individual is calculated. Based on the preset number of elites, the selection ratio is allocated according to the fitness value, and the elite individuals are determined by random proportion.
[0043] Through the crossover operation, according to the preset crossover number, the planning schemes of any pair of elite individuals are partially exchanged to generate new individuals; through the mutation operation, according to the preset mutation number, any other individuals are selected to randomly explore in the solution space to generate mutant individuals, and the elite individuals, the new individuals and the mutant individuals are formed into a first type of new population;
[0044] Based on the preset migration interval, after the corresponding rounds of iterations, according to the preset migration ratio, the corresponding number of individuals with the largest fitness values are selected from each population as migration individuals, and the individuals are exchanged and migrated among all populations to generate the second type of new population; according to the preset alliance size, the corresponding number of populations are merged to generate the third type of new population;
[0045] The first type of new population, the second type of new population and the third type of new population together form a new community, and the iteration continues;
[0046] Until the preset number of iterations is reached, the individual with the largest fitness value is selected from the final iterative community to obtain the optimal planning solution.
[0047] According to a second aspect of the embodiments of the present invention,
[0048] Provided is a system resource requirement planning system based on artificial intelligence, comprising:
[0049] The first unit is used to obtain system resource data, construct a resource attribute heterogeneous graph, generate a resource node sequence set through multiple rounds of random walks based on the resource attribute heterogeneous graph, learn the node neighborhood relationship through a training graph embedding model, construct a candidate feature matrix, and generate an optimal feature subset based on the comprehensive feature score of each candidate feature;
[0050] The second unit is used to input the system resource time series data to be predicted into the resource demand prediction network, generate mixed Gaussian distribution parameters by capturing the dependency relationship of each time step, determine the conditional probability distribution, and obtain the resource demand prediction result;
[0051] The third unit is used to randomly generate multiple planning schemes according to the resource demand prediction results and the system resources determined based on the optimal feature subset, construct an iterative community, input the iterative community into the resource allocation algorithm, generate the optimal solution, and determine the optimal planning scheme.
[0052] According to a third aspect of the embodiments of the present invention,
[0053] An electronic device is provided, comprising:
[0054] processor;
[0055] a memory for storing processor-executable instructions;
[0056] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.
[0057] A fourth aspect of the embodiments of the present invention is:
[0058] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.
[0059] In an embodiment of the present invention, by constructing a heterogeneous graph of resource attributes, system resources and their complex interrelationships are represented in a structured manner, multiple relationships between resources are clearly described, the heterogeneity of resources is captured, and different attributes and relationships of different types of resources are reflected; multiple rounds of random walks can effectively capture local and global information of nodes in the graph, and by considering neighboring nodes and their relationships, local structural characteristics of nodes can be captured, and the global structural characteristics of the graph can be captured through the exploration and return mechanism during the random walk process; by generating mixed Gaussian distribution parameters, the model can flexibly describe complex nonlinear relationships and capture multimodal characteristics of resource demand. The modeling method allows probabilistic description of multiple possible situations of resource demand, thereby improving the flexibility and accuracy of prediction; the generated mixed Gaussian distribution parameters can be used to determine conditional probability distribution, so that the prediction results have a probabilistic interpretation, which can not only give the most likely prediction value, but also provide a range of uncertainty in the prediction, helping decision makers to better evaluate and respond to potential risks and fluctuations; by introducing a multi-population co-evolution mechanism, collaborative optimization within and between populations is achieved, and the diversity and complementarity of different populations are fully utilized to improve the efficiency and quality of resource planning optimization. BRIEF DESCRIPTION OF THE DRAWINGS
[0060] Figure 1 A flowchart of a method for planning system resource requirements based on artificial intelligence according to an embodiment of the present invention;
[0061] Figure 2 It is a structural diagram of a system resource requirement planning system based on artificial intelligence according to an embodiment of the present invention. DETAILED DESCRIPTION
[0062] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0063] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.
[0064] Figure 1 FIG. 1 is a flow chart of a method for planning system resource requirements based on artificial intelligence according to an embodiment of the present invention. Figure 1 As shown, the method includes:
[0065] S101. Obtain system resource data, construct a resource attribute heterogeneous graph, generate a resource node sequence set through multiple rounds of random walks based on the resource attribute heterogeneous graph, learn the node neighborhood relationship through the training graph embedding model, construct a candidate feature matrix, and generate an optimal feature subset based on the comprehensive feature score of each candidate feature;
[0066] The resource attribute heterogeneous graph is a graph data structure in which nodes represent different types of resources or entities, and edges represent multiple relationships between these resources or entities;
[0067] The resource attribute heterogeneous graph contains the following elements:
[0068] Nodes represent different types of system resource entities or attributes. Entities include servers, virtual machines, storage devices, network devices, applications, etc. Attributes include CPU, memory, operating system, allocated resource amount, etc.
[0069] Edges represent the relationship between system resource entities, or the inclusion relationship between entities and attributes, and can be of various types, such as server-virtual machine, virtual machine-storage device, server-network device, etc. Edges can carry weights to indicate the strength or importance of the relationship; the edge between a server and a virtual machine can represent the hosting relationship of the virtual machine, and the weight carried can indicate the degree of hosting load; the edge between a virtual machine and a storage device can represent a storage connection, and the weight carried can indicate the connection bandwidth or usage;
[0070] Resource attribute heterogeneous graphs can fully express complex resource and relationship networks through their heterogeneity, that is, different types of nodes and edges.
[0071] In this embodiment, by constructing a heterogeneous graph of resource attributes, system resources and their complex relationships are represented in a structured manner, various relationships between resources are clearly described, the heterogeneity of resources is captured, and different attributes and relationships of different types of resources are reflected; multiple rounds of random walks can effectively capture local and global information of nodes in the graph, and by considering neighboring nodes and their relationships, the local structural characteristics of nodes can be captured. Through the exploration and return mechanism during the random walk, the global structural characteristics of the graph can be captured; by training the graph embedding model, the neighborhood relationship of nodes can be learned, and the high-dimensional graph structure information can be embedded into the low-dimensional vector space. Similar nodes have closer distances in the embedded space, reducing the feature dimension, making subsequent machine learning model training more efficient; representing various attributes and relationships of nodes as feature vectors helps to comprehensively describe the characteristics of nodes, comprehensively consider node attributes, neighbor attributes and graph structure characteristics, and enrich the feature set.
[0072] In an optional embodiment, system resource data is obtained, a resource attribute heterogeneous graph is constructed, and based on the resource attribute heterogeneous graph, a resource node sequence set is generated through multiple rounds of random walks, and the node neighborhood relationship is learned through a training graph embedding model to construct a candidate feature matrix. Based on the comprehensive feature score of each candidate feature, an optimal feature subset is generated, including:
[0073] Obtain system resource data, determine resource entities and resource attributes, use resource entities and resource attributes as resource nodes, and use association weights between resource nodes as connecting edges to construct a resource heterogeneous graph;
[0074] Randomly select a central node from the resource heterogeneous graph, determine the context node through random walk, and generate a resource node sequence;
[0075] After multiple rounds of random walks, multiple resource node sequences are generated to form a resource node sequence set;
[0076] Taking the resource node sequence set as data samples, the central node as input, and the context node as output, based on the pre-built graph embedding model, learning the node neighborhood relationship, generating a low-dimensional dense vector representation corresponding to each node, and combining them into a candidate feature matrix;
[0077] Based on each column of feature vectors in the candidate feature matrix, the feature comprehensive scores are calculated and sorted, and according to the preset optimal number of features, the corresponding number of candidate features are selected to generate the optimal feature subset.
[0078] By accessing the system's resource management interface, database or log file, etc., the resource data in the system is obtained. The resource data includes various resource entities in the system and attribute information describing the characteristics of the resource entities. Different types of resource entities are identified from the obtained resource data. Resource entities can be physical devices in the system, such as servers, storage devices, network devices, etc.; they can also be virtual resources, such as virtual machines, containers, virtual networks, etc.; they can also be services that provide specific functions, such as database services, cache services, load balancing services, etc. Each resource entity has a unique identifier to distinguish different entities; for each identified resource entity, its corresponding attribute information is extracted. Resource attributes describe the various characteristics of resource entities. Different types of resource entities may have different sets of attributes; based on the determined resource entities and resource attributes, nodes in the resource heterogeneous graph are generated. Each resource entity corresponds to a node. The resource attribute is an independent node connected to the resource entity node, indicating that the entity has the attribute. The association between resource entities is analyzed, and the connection edges between resource nodes are determined;
[0079] According to the type and strength of the association between resource entities, determine the edges connecting resource nodes and the weights of the edges. The weight of the edge can reflect the closeness or importance of the association relationship. According to the generated resource nodes and the determined node association relationship, a heterogeneous graph representing system resources is constructed. The heterogeneous graph contains different types of resource nodes and weighted edges connecting different nodes. The heterogeneous graph comprehensively describes the composition, structure and association of resources in the system, providing a basis for subsequent resource analysis and optimization.
[0080] A node is randomly selected from the constructed resource heterogeneous graph as the central node, and sampling is performed on the heterogeneous graph by random walk. Random walk means starting from the central node, randomly selecting the next node as the walking target with a certain probability according to the weight of the connection edge between nodes, until the preset number of walking steps is reached or specific conditions are met. During the random walk, the central node and the visited nodes are recorded to generate a resource node sequence, in which the central node is located at the center of the sequence, and the visited nodes are used as the context nodes of the central node.
[0081] Through multiple rounds of random walks, multiple resource node sequences are generated to form a set of resource node sequences. Each sequence represents a combination of a central node and its context nodes, reflecting the local structural relationship between nodes in the heterogeneous graph.
[0082] The resource node sequence set is used as the data sample, the central node in each sequence is used as the input, and the corresponding context node is used as the output. The pre-built graph embedding model is used to learn the node representation. The graph embedding model preferably uses the Word2Vec model. By optimizing the model parameters, the neighborhood relationship between nodes is learned, and each node is mapped to a low-dimensional dense vector space. The learned node vector representation retains the structural information and semantic information of the node in the heterogeneous graph. The distance between node vectors can reflect the degree of association between nodes. The vector representations of all nodes are combined into a candidate feature matrix, and each column corresponds to the feature vector of a node.
[0083] Calculate the feature comprehensive score for each column of feature vectors in the candidate feature matrix. The feature comprehensive score takes into account multiple factors such as the importance, discrimination, and stability of the feature, and preferably uses the information gain feature evaluation indicator; sort the candidate features according to the feature comprehensive score, and select the top N features with the highest scores as the optimal feature subset, where N is the preset number of optimal features. The features in the optimal feature subset can retain the structure and semantic information of the resource heterogeneous graph to the greatest extent, and have good discrimination ability and stability, and can be used for subsequent resource analysis, prediction and optimization tasks;
[0084] In this embodiment, a resource heterogeneous graph is constructed according to resource entities and attribute information, wherein each resource entity corresponds to a node, the connecting edges between nodes represent the association relationship between resource entities, and the weight of the edges reflects the strength or importance of the association relationship; the heterogeneous graph comprehensively depicts the composition, structure and association of resources in the system, and provides a basis for subsequent resource analysis and optimization; through multiple rounds of random walks, sampling is performed on the heterogeneous graph to generate multiple sets of resource node sequences, which reflect the local structural relationship between nodes and can effectively capture the local information of the nodes; each resource node sequence contains a combination of a central node and its context nodes, which can be used for subsequent graph embedding model training; through the pre-constructed graph embedding model, the nodes are represented and learned, and each node is mapped to a low-dimensional dense vector space, retaining the structural information and semantic information of the node in the heterogeneous graph; the vector representations of all nodes are combined into a candidate feature matrix, and each column corresponds to a feature vector of a node; the feature vectors in each column of the candidate feature matrix are scored for comprehensive features, taking into account multiple factors such as the importance, discrimination, and stability of the features, and preferably using the information gain feature evaluation indicator.
[0085] In an optional embodiment, randomly selecting a central node from the resource heterogeneous graph, determining a context node through random walk, and generating a resource node sequence includes:
[0086] From the resource heterogeneous graph, randomly select the resource node corresponding to the resource entity as the starting node, take the starting node as the current node, and start random walk:
[0087] Based on the current node, calculate the node degree, determine the return factor, and calculate the neighborhood clustering coefficient to determine the exploration factor. Combine the weights of all the edges connected to the current node to determine the transition probability distribution corresponding to the neighboring node. Based on the transition probability distribution, set a random roulette wheel to determine the next node and add it to the node sequence.
[0088] The next node is set as the current node, and the random walk is repeated until the length of the node sequence reaches a preset sequence length, thereby generating a resource node sequence.
[0089] The node degree specifically refers to the number of edges that a node directly connects to other nodes. For an undirected graph, the node degree is the number of neighbor nodes of the node.
[0090] The return factor specifically refers to the probability of controlling the return to the last visited node during the random walk, and is dynamically adjusted according to the degree of the current node. For nodes with larger degrees, the value of the return factor increases accordingly to encourage the walk to return to the previous node; for nodes with smaller degrees, the value of the return factor decreases accordingly to encourage the walk to explore new nodes.
[0091] The clustering coefficient specifically refers to an indicator used to measure the degree of connection between neighbors of a node, reflecting the degree of connection between nodes in the subgraph where the node is located, that is, the probability that there is a connection between neighbor nodes of the node. For a node, the higher the clustering coefficient, the closer the connection between its neighbor nodes.
[0092] The exploration factor specifically refers to the factor used to control the breadth and depth of exploring the neighborhood of the current node during the walk, and is dynamically adjusted according to the clustering coefficient of the neighborhood of the current node. For nodes with a higher clustering coefficient, the value of the exploration factor is reduced accordingly, encouraging the walk to conduct in-depth exploration within the current community; for nodes with a lower clustering coefficient, the value of the exploration factor is increased accordingly, encouraging the walk to jump out of the current community and explore a broader graph structure.
[0093] Randomly select a node corresponding to a resource entity from the resource heterogeneous graph as the starting node. The starting node can be selected by uniform random sampling;
[0094] Set the selected starting node as the current node, initialize an empty node sequence to record the nodes visited during the random walk, and set the parameters of the random walk, such as the maximum number of walk steps, sequence length, etc.
[0095] For the current node, calculate the node degree, that is, the number of edges directly connected to the current node. The node degree reflects the importance and influence of the node. According to the degree of the current node, determine a return factor to control the probability of returning to the visited node during the random walk. Preferably, the return factor is set to a value proportional to the degree. The larger the degree, the higher the probability of returning to the visited node, so as to ensure repeated visits to important nodes during the walk;
[0096] For the current node, calculate the clustering coefficient between the nodes in its neighborhood. The clustering coefficient reflects the closeness of the node neighborhood. The higher the clustering coefficient, the denser the connection between the nodes in the neighborhood. According to the clustering coefficient of the current node neighborhood, determine an exploration factor to control the probability of exploring new nodes during the random walk. The exploration factor can be set to a value inversely proportional to the clustering coefficient. The higher the clustering coefficient, the lower the probability of exploring new nodes, so as to encourage the walk process to conduct in-depth exploration in the tightly connected local community.
[0097] Based on the weights of all edges directly connected to the current node, the transition probability of each neighbor node being selected as the next wandering target is calculated. The transition probability distribution preferably uses normalized edge weights, that is, the edge weight of each neighbor node is divided by the sum of the edge weights of all neighbor nodes to obtain the transition probability of the corresponding neighbor node. The transition probability distribution reflects the relative likelihood of different neighbor nodes being visited. The neighbor node with a larger weight has a higher probability of being selected.
[0098] According to the calculated transfer probability distribution, a random roulette wheel is set to determine the next target node for the walk. The transfer probability distribution is mapped to an interval between 0 and 1, and then a random number is generated in the corresponding interval. According to the sub-interval where the random number falls, the corresponding neighbor node is determined as the next walk target. The neighbor node with a higher transfer probability has a greater chance of being selected, while the neighbor node with a lower transfer probability has a smaller chance of being selected.
[0099] The next node selected by the random wheel is added to the node sequence, indicating that the corresponding node has been visited during the random walk. The next node is set as the new current node, and the random walk continues. The random walk is repeated until the length of the node sequence reaches the preset sequence length, that is, the number of nodes visited during the walk reaches the predetermined target value;
[0100] The conditions for stopping the random walk also include: reaching a preset maximum number of walk steps, that is, the total number of random walk steps exceeds the set upper limit; the distribution of visited nodes during the random walk tends to be stable, and the new walk no longer significantly changes the statistical characteristics of the node sequence.
[0101] When the termination condition of the random walk is met, the walk process is stopped and the recorded node sequence is used as a resource node sequence. The corresponding resource node sequence reflects the order and frequency of accessing nodes during the random walk process, and contains the local structure and semantic information of the resource heterogeneous graph.
[0102] In this embodiment, by calculating the node degree and setting the return factor, important nodes can be visited more frequently during the random walk, ensuring full attention to key resource entities; by using the clustering coefficient and the exploration factor, deep exploration can be controlled within a tightly connected local community to capture detailed information of the local structure, while jumping out of the local community in a timely manner to ensure comprehensive coverage of the global structure; the transfer probability of neighbor nodes is calculated based on edge weights to ensure that neighbor nodes with close relationships with the current node are more likely to be visited during the walk, reflecting the intensity of interaction between resources in the actual system; the next walk node is selected by the random roulette method, so that the transfer process has both randomness and controllability, enhancing the diversity and representativeness of the walk sequence.
[0103] In an optional embodiment, based on the current node, the node degree is calculated, the return factor is determined, and the clustering coefficient of the neighborhood is calculated, the exploration factor is determined, and the weights of all connecting edges connected to the current node are combined to determine the transition probability distribution corresponding to the neighboring node, including:
[0104] The transition probability from the current node to the neighboring node is as follows:
[0105] ;
[0106] in, u Representation Node u , v Represents another node v , P ( u | v ) indicates a slave node v To Node u The transition probability, β represents the balance factor, deg( v ) represents a node v The node degree of N ( v ) represents a node v The set of neighboring nodes of x Represents a node in the neighborhood node set x , C u Representation Node u The clustering coefficient of the neighborhood, C x Representation Node x The clustering coefficient of the neighborhood, ω vu Representation Node v and nodes u The weight of the edge connecting them, y Represents a node in the neighborhood node set y , C y Representation Node y The clustering coefficient of the neighborhood, ω vx Representation Node v and nodes x The weight of the connecting edge.
[0107] Indicates the return factor, node v The influence of degree on the transfer probability, the larger the degree, the higher the probability of return; represents the exploration factor, node u The influence of the neighborhood clustering coefficient on the transition probability. The higher the clustering coefficient, the lower the probability of exploring new nodes. The normalized denominator is the current node v All neighbor nodes x The sum of the product of the transfer factor and the weight of the connecting edge ensures that the sum of the transfer probabilities is 1, so that it meets the requirements of the probability distribution;
[0108] From the current node v To neighbor node uThe transfer probability is determined by the following factors: the node degree reflects the importance of the node, the return factor is determined by the node degree, the larger the degree, the higher the return probability; the neighborhood clustering coefficient reflects the tightness of the connection between nodes in the neighborhood, the exploration factor is calculated based on the clustering coefficient, the higher the clustering coefficient, the lower the probability of exploring new nodes; the weight of the connection edge reflects the strength of the relationship between nodes, the larger the weight, the higher the probability of transferring to the node;
[0109] By combining the return factor, exploration factor and edge weight, the edge weight from the current node is calculated. v To neighbor node u The transition probability of the neighboring nodes is normalized to ensure that the sum of the transition probabilities of all neighboring nodes is 1.
[0110] According to the formula, the node degree reflects the importance of the node. By introducing the degree into the calculation of the transition probability, it is possible to ensure that appropriate attention is paid to important nodes during the random walk and avoid missing key nodes. The clustering coefficient reflects the compactness of the local neighborhood of the node. By considering the clustering coefficient, in-depth exploration can be performed in the tightly connected local community to capture local structural information. By balancing the return factor and the exploration factor, the probability of returning to the visited nodes and exploring new nodes can be adaptively adjusted during the random walk. The adaptive mechanism enables the random walk strategy to be dynamically adjusted according to the local characteristics of the nodes, thereby improving the flexibility and accuracy of the walk. It is proportional to the node degree. The larger the degree, the higher the probability of returning to the visited nodes, ensuring that important nodes receive more attention. It is inversely proportional to the clustering coefficient. The higher the clustering coefficient, the lower the probability of exploring new nodes, encouraging in-depth exploration in the tightly connected local community. Combined with the introduction of the weight of the connecting edge, the transition probability can be adjusted according to the strength of the relationship between the nodes, so that the edges with larger weights are more likely to be selected, thereby increasing the attention paid to important edges during the walk and enhancing the accurate reflection of the walk strategy on the graph structure.
[0111] S102. Input the system resource time series data to be predicted into the resource demand prediction network, generate mixed Gaussian distribution parameters by capturing the dependency of each time step, determine the conditional probability distribution, and obtain the resource demand prediction result;
[0112] The mixed Gaussian distribution specifically refers to a probability distribution model, which is composed of a weighted combination of multiple Gaussian distributions. Each Gaussian distribution is called a component in the mixed model, and each component has its corresponding weight; the mixed Gaussian distribution parameters are: weight, mean, and standard deviation, where the weight reflects the contribution of each Gaussian distribution to the overall distribution. A larger weight indicates that the Gaussian distribution has a greater impact on the final mixed distribution; the mean reflects the central tendency of the data described by the component and is the main value predicted by the component; the standard deviation reflects the volatility of the data described by the component. The larger the standard deviation, the looser the data distribution and the greater the volatility.
[0113] In this embodiment, by generating mixed Gaussian distribution parameters, the model can flexibly describe complex nonlinear relationships and capture the multimodal characteristics of resource demand. The modeling method allows for a probabilistic description of various possible situations of resource demand, thereby improving the flexibility and accuracy of prediction; the generated mixed Gaussian distribution parameters can be used to determine the conditional probability distribution, so that the prediction results have a probabilistic interpretation, which can not only give the most likely prediction value, but also provide a range of uncertainty in the prediction, helping decision makers to better assess and respond to potential risks and fluctuations; by capturing the dependencies of time series and generating mixed Gaussian distribution parameters, the model can comprehensively consider various factors affecting resource demand and improve the overall prediction accuracy. This is of great significance to resource planning and management, and can effectively avoid waste or shortage of resources; the model can dynamically generate mixed Gaussian distribution parameters based on the input time series data to adapt to changes in resource demand in different time periods and different situations. Such adaptive capabilities enable the model to still provide reliable predictions in the face of emergencies or demand fluctuations.
[0114] In an optional embodiment, the system resource time series data to be predicted is input into the trained optimal resource demand prediction network, and the mixed Gaussian distribution parameters are generated by capturing the dependency relationship of each time step, and the conditional probability distribution is determined to obtain the resource demand prediction results including:
[0115] Acquire time series data of system resources, preprocess the time series data, divide the time series data into multiple groups based on a preset sliding window size, determine an input sequence and an output sequence, and generate a training sample set;
[0116] Constructing a resource demand prediction network, including an input layer, at least one recursive hidden layer and an output layer, wherein the input layer receives the input sequence; the recursive hidden layer is constructed based on a long short-term memory network, and controls the information transmission of the input sequence through a gating mechanism to capture long-term dependencies; the output layer includes a plurality of fully connected layer structures, and generates mixed Gaussian distribution parameters through direct output, exponential function transformation output and activation function normalized output respectively;
[0117] Inputting the training sample set into the resource demand prediction network, training based on the gradient descent method, iteratively adjusting the model parameters by maximizing the loss function until a preset number of iterations is reached to generate an optimal resource demand prediction network;
[0118] The time series data to be predicted is input into the optimal resource demand prediction network to obtain mixed Gaussian distribution parameters, determine the conditional probability distribution, and determine the resource demand prediction result by calculating the expectation of the conditional probability distribution.
[0119] Based on the preset sliding window size, the preprocessed time series data is segmented. The size of the sliding window determines the input sequence length of each training sample. The sliding window moves on the time series, extracts the input sequence and the corresponding output sequence in turn, and forms multiple sets of training samples, where the input sequence is the time series fragment within the sliding window, and the output sequence is the resource demand value of several time steps after the end position of the sliding window. Repeat the sliding window segmentation until the end of the time series to generate a complete set of training samples;
[0120] A deep learning model, called a resource demand prediction network, is constructed to learn long-term dependencies in time series data and predict future resource requirements. The resource demand prediction network includes an input layer, at least one recursive hidden layer, and an output layer. The input layer receives the input sequence in the training sample and passes it to the subsequent recursive hidden layer.
[0121] The recursive hidden layer adopts the long short-term memory network structure, and controls the transmission and update of information through the gating mechanism. The long short-term memory network contains input gate, forget gate and output gate, which respectively control the addition of new information, the forgetting of historical information and the output of the current state. Through the gating mechanism, the long short-term memory network can selectively retain or forget previous information and capture long-term dependencies in the time series. The recursive hidden layer contains multiple long short-term memory network layers to extract higher-level feature representations;
[0122] The output layer includes multiple fully connected layers, which transform and integrate the output of the recursive hidden layer to generate the final prediction result. The structure of the output layer is designed to generate the parameters of the mixed Gaussian distribution, including the mean, variance and mixing coefficient. In order to obtain these parameters, the output layer adopts three different output methods: directly output the mean, output the non-negative variance through exponential function transformation, and normalize the output mixing coefficient through the activation function, preferably the softmax function; the output layer generates a set of parameters of the mixed Gaussian distribution, which are used to describe the probability distribution of future resource demand;
[0123] The generated training sample set is input into the constructed resource demand prediction network to start the model training process. The gradient descent method is used to optimize the model parameters by maximizing the loss function.
[0124] The loss function is designed as a negative log-likelihood function, which minimizes the difference between the predicted probability distribution and the true resource demand value. The predicted mixed Gaussian distribution parameters are calculated through forward propagation, and the negative log-likelihood loss is calculated based on the true resource demand value.
[0125] In each training iteration, the gradient of the loss function to the model parameters is calculated through the back propagation algorithm, and the model parameters are updated using the gradient descent method. The training iteration is repeated until the preset number of iterations is reached to obtain the optimal resource demand prediction network model;
[0126] The optimal resource demand prediction network model obtained through training is used to predict future resource demand.
[0127] The time series data to be predicted is input into the optimal resource demand prediction network. Through forward propagation calculation, the mixed Gaussian distribution parameters generated by the output layer are obtained, including the mean, variance and mixing coefficient. According to the mixed Gaussian distribution parameters, the conditional probability distribution is determined. The conditional probability distribution describes the probability distribution of future resource demand under a given input sequence. The expected value of the conditional probability distribution is calculated to obtain the predicted value of future resource demand. The expected value represents the average value or the most likely value of the resource demand under the probability distribution.
[0128] In this embodiment, the time series data is divided into training samples by a preset sliding window size, which can capture local and global features in the time series, so that the model can better learn patterns and trends in the time series; effectively capture long-term dependencies in time series data and improve prediction accuracy; multi-layer long short-term memory networks can extract high-level features of time series data, capture more complex patterns and changes, and improve the model's ability to predict future resource demand; the mixed Gaussian distribution parameters generated by the output layer, including mean, variance and mixing coefficient, can perform probabilistic modeling on future resource demand, and the prediction method of the probability distribution not only provides the most likely prediction value, but also describes the uncertainty of the prediction results, providing richer information; the model parameters are optimized by the gradient descent method and the back propagation algorithm, so that the model is continuously improved during the training process, and finally the optimal resource demand prediction network model is obtained to ensure the accuracy and reliability of the prediction results.
[0129] In an optional embodiment, by maximizing the loss function and iteratively adjusting the model parameters until a preset number of iterations is reached, generating an optimal resource demand prediction network includes:
[0130] The loss function has the following formula:
[0131] ;
[0132] in, L represents the loss function, i represents the Gaussian component ordinal number, M represents the total number of mixed distributions, π i Indicates iThe mixing coefficients of Gaussian components, p represents the variables in the input sequence, q represents the variables in the output sequence, μ i Indicates i The mean of the Gaussian components, σ i Indicates i The standard deviation of the Gaussian components.
[0133] According to the formula, using mixed Gaussian distribution to describe future resource demand can capture the multimodal characteristics of time series data and better adapt to complex and changeable demand patterns; by maximizing the probability density of observed data under the mixed Gaussian distribution, the model can generate probabilistic prediction results, which not only gives the most likely prediction value, but also provides uncertainty information of the prediction results, thereby enhancing the reliability of decision-making; the negative log-likelihood loss function is used to optimize the model so that the mixed Gaussian distribution parameters (mean, variance, mixing coefficient) generated by the model match the real data as accurately as possible, thereby improving the accuracy of the prediction; the loss function can effectively avoid numerical instability, especially when dealing with small probability density, to ensure the stability and reliability of the training process; by modeling the relationship between the input sequence and the output sequence, the loss function can comprehensively capture the dependency of time series data and enhance the model's ability to respond to future changes in resource demand.
[0134] S103. According to the resource demand prediction results and the system resources determined based on the optimal feature subset, a plurality of planning schemes are randomly generated, an iterative community is constructed, the iterative community is input into the resource allocation algorithm, an optimal solution is generated, and an optimal planning scheme is determined.
[0135] In this embodiment, an iterative community is constructed, and through iterative optimization of the community, each planning scheme is gradually improved to make it more in line with resource requirements and optimization goals, thereby improving the overall optimization efficiency. The characteristics of the swarm intelligence algorithm are utilized to find a better solution through collaborative evolution and information sharing within the community. The resource allocation algorithm can accurately optimize the resource allocation plan under complex constraints to ensure maximum resource utilization and optimal system performance. The algorithm can be dynamically adjusted according to real-time changing resource requirements and system status to improve the flexibility and responsiveness of the system.
[0136] In an optional embodiment, the iterative process of the resource allocation algorithm includes:
[0137] Taking the planning scheme as an individual, initialize multiple populations and construct an iterative community;
[0138] Based on the preset fitness function, the fitness value of each individual is calculated. Based on the preset number of elites, the selection ratio is allocated according to the fitness value, and the elite individuals are determined by random proportion.
[0139] Through the crossover operation, according to the preset crossover number, the planning schemes of any pair of elite individuals are partially exchanged to generate new individuals; through the mutation operation, according to the preset mutation number, any other individuals are selected to randomly explore in the solution space to generate mutant individuals, and the elite individuals, the new individuals and the mutant individuals are formed into a first type of new population;
[0140] Based on the preset migration interval, after the corresponding rounds of iterations, according to the preset migration ratio, the corresponding number of individuals with the largest fitness values are selected from each population as migration individuals, and the individuals are exchanged and migrated among all populations to generate the second type of new population; according to the preset alliance size, the corresponding number of populations are merged to generate the third type of new population;
[0141] The first type of new population, the second type of new population and the third type of new population together form a new community, and the iteration continues;
[0142] Until the preset number of iterations is reached, the individual with the largest fitness value is selected from the final iterative community to obtain the optimal planning solution.
[0143] Represent resource planning schemes as individuals, each of which represents a possible resource allocation and scheduling scheme. Initialize multiple populations, each of which contains a certain number of individuals. The number of populations and the number of individuals in each population can be set according to the problem scale and computing resources. The initialized populations are organized into iterative communities as the starting point of evolutionary optimization.
[0144] For each population in the iteration population, do the following:
[0145] Fitness evaluation, based on the preset fitness function, calculates the fitness value of each individual in the population; the fitness function measures the quality of the planning scheme corresponding to the individual, taking into account factors such as resource utilization, load balancing, and performance indicators. The larger the fitness value, the better the planning scheme corresponding to the individual;
[0146] Elite selection: According to the preset number of elites, the selection ratio is determined according to the individual fitness value. The higher the fitness value of the individual, the greater the probability of being selected as an elite individual. The elite individuals are selected from the population in a random manner as representatives of high-quality solutions.
[0147] Crossover operation: According to the preset number of crossovers, elite individuals are randomly selected to form crossover pairs. For each crossover pair, some of the planning schemes of individuals are randomly selected to be exchanged to generate new individuals. The crossover operation explores new solution spaces by combining some of the solutions of high-quality solutions to generate potentially better planning schemes.
[0148] The mutation operation randomly selects individuals from the population for mutation according to the preset number of mutations. For each mutated individual, it randomly explores the solution space, modifies or reconstructs some of the individual's planning schemes, and generates new mutated individuals. The mutation operation introduces randomness to jump out of the local optimum and explore a broader solution space.
[0149] The new population is generated by combining elite individuals, new individuals generated by crossover operations, and mutated individuals generated by mutation operations to form a new population; the new population inherits the characteristics of high-quality solutions and introduces new diversity, providing a better foundation for subsequent evolution.
[0150] The following coordinated operations are performed between populations in the iterative community:
[0151] Migration operation: Based on the preset migration interval, after a certain number of iterations, the migration operation is triggered. According to the preset migration ratio, the individual with the largest fitness value is selected from each population as the migration individual, and the migration individuals are exchanged between different populations, so that high-quality solutions can be spread among populations, promoting information exchange and co-evolution between populations.
[0152] Alliance operation: according to the preset alliance scale, some populations in the iteration community are merged to form a new alliance population. The alliance population brings together high-quality solutions from different populations, and accelerates the evolution process and improves optimization efficiency by sharing information and resources.
[0153] Repeat the operations within the population and the collaborative operations between populations, and continuously update the iterative community until the preset number of iterations is reached. The number of iterations is set according to the complexity of the problem and the computing resources to balance the optimization effect and computing cost;
[0154] After the iteration is terminated, the individual with the largest fitness value is selected from the final iterative community as the optimal solution. The planning scheme corresponding to the optimal solution is the optimal resource planning scheme found, which can guide the actual resource allocation and scheduling process.
[0155] In this embodiment, by introducing a multi-population co-evolution mechanism, collaborative optimization within and between populations is achieved, the diversity and complementarity of different populations are fully utilized, and the efficiency and quality of resource planning optimization are improved; within the population, through fitness evaluation, elite selection, crossover operation and mutation operation, the retention of high-quality solutions and the exploration of new solutions are achieved, the utilization and exploration of search are balanced, premature convergence to the local optimum is avoided, and the global search capability of the algorithm is improved; between populations, through migration operations and alliance operations, information exchange and resource sharing between populations are achieved, the propagation and integration of high-quality solutions between different populations are promoted, the evolution process is accelerated, and the optimization efficiency is improved; through the design of fitness function, multiple optimization objectives such as resource utilization, load balancing, performance indicators, etc. are comprehensively considered to achieve a comprehensive evaluation and optimization of resource planning schemes, and the optimal solution obtained can achieve efficient allocation and scheduling of resources in multiple dimensions.
[0156] Figure 2 FIG. 1 is a schematic diagram of the structure of a system resource requirement planning system based on artificial intelligence according to an embodiment of the present invention. Figure 2 As shown, the system comprises:
[0157] The first unit is used to obtain system resource data, construct a resource attribute heterogeneous graph, generate a resource node sequence set through multiple rounds of random walks based on the resource attribute heterogeneous graph, learn the node neighborhood relationship through a training graph embedding model, construct a candidate feature matrix, and generate an optimal feature subset based on the comprehensive feature score of each candidate feature;
[0158] The second unit is used to input the system resource time series data to be predicted into the resource demand prediction network, generate mixed Gaussian distribution parameters by capturing the dependency relationship of each time step, determine the conditional probability distribution, and obtain the resource demand prediction result;
[0159] The third unit is used to randomly generate multiple planning schemes according to the resource demand prediction results and the system resources determined based on the optimal feature subset, construct an iterative community, input the iterative community into the resource allocation algorithm, generate the optimal solution, and determine the optimal planning scheme.
[0160] According to a third aspect of the embodiments of the present invention,
[0161] An electronic device is provided, comprising:
[0162] processor;
[0163] a memory for storing processor-executable instructions;
[0164] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.
[0165] A fourth aspect of the embodiments of the present invention is:
[0166] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.
[0167] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.
[0168] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A system resource requirement planning method based on artificial intelligence, characterized in that: include: Acquire system resource data, construct a resource attribute heterogeneous graph, generate a resource node sequence set through multiple rounds of random walks based on the resource attribute heterogeneous graph, learn the node neighborhood relationship through a training graph embedding model, construct a candidate feature matrix, and generate an optimal feature subset based on the comprehensive feature score of each candidate feature; The system resource time series data to be predicted is input into the resource demand prediction network, and the mixed Gaussian distribution parameters are generated by capturing the dependency relationship of each time step, and the conditional probability distribution is determined to obtain the resource demand prediction result; According to the resource demand prediction result and the system resources determined based on the optimal feature subset, a plurality of planning schemes are randomly generated, an iterative community is constructed, the iterative community is input into the resource allocation algorithm, an optimal solution is generated, and an optimal planning scheme is determined; Obtain system resource data, build a resource attribute heterogeneous graph, generate a resource node sequence set through multiple rounds of random walks based on the resource attribute heterogeneous graph, learn the node neighborhood relationship through the training graph embedding model, build a candidate feature matrix, and generate the optimal feature subset based on the comprehensive feature score of each candidate feature, including: Obtain system resource data, determine resource entities and resource attributes, use resource entities and resource attributes as resource nodes, and use association weights between resource nodes as connecting edges to construct a resource heterogeneous graph; Randomly select a central node from the resource heterogeneous graph, determine the context node through random walk, and generate a resource node sequence; After multiple rounds of random walks, multiple resource node sequences are generated to form a resource node sequence set; Taking the resource node sequence set as data samples, the central node as input, and the context node as output, based on the pre-built graph embedding model, learning the node neighborhood relationship, generating a low-dimensional dense vector representation corresponding to each node, and combining them into a candidate feature matrix; Based on each column of feature vectors in the candidate feature matrix, the feature comprehensive score is calculated and sorted, and according to the preset optimal feature number, a corresponding number of candidate features are selected to generate an optimal feature subset; From the resource heterogeneous graph, randomly selecting a central node, determining a context node through random walk, and generating a resource node sequence includes: From the resource heterogeneous graph, randomly select the resource node corresponding to the resource entity as the starting node, take the starting node as the current node, and start random walking: Based on the current node, calculate the node degree, determine the return factor, and calculate the neighborhood clustering coefficient to determine the exploration factor. Combine the weights of all the edges connected to the current node to determine the transition probability distribution corresponding to the neighboring node. Based on the transition probability distribution, set a random roulette wheel to determine the next node and add it to the node sequence. The next node is set as the current node, and the random walk is repeated until the length of the node sequence reaches a preset sequence length, thereby generating a resource node sequence; The system resource time series data to be predicted is input into the trained optimal resource demand prediction network. By capturing the dependency of each time step, the mixed Gaussian distribution parameters are generated, and the conditional probability distribution is determined. The resource demand prediction results include: Acquire time series data of system resources, preprocess the time series data, divide the time series data into multiple groups based on a preset sliding window size, determine an input sequence and an output sequence, and generate a training sample set; Constructing a resource demand prediction network, including an input layer, at least one recursive hidden layer and an output layer, wherein the input layer receives the input sequence; the recursive hidden layer is constructed based on a long short-term memory network, and controls the information transmission of the input sequence through a gating mechanism to capture long-term dependencies; the output layer includes a plurality of fully connected layer structures, and generates mixed Gaussian distribution parameters through direct output, exponential function transformation output and activation function normalized output respectively; Inputting the training sample set into the resource demand prediction network, training based on the gradient descent method, iteratively adjusting the model parameters by maximizing the loss function until a preset number of iterations is reached to generate an optimal resource demand prediction network; Input the time series data to be predicted into the optimal resource demand prediction network, obtain mixed Gaussian distribution parameters, determine the conditional probability distribution, and determine the resource demand prediction result by calculating the expectation of the conditional probability distribution; The iterative process of the resource allocation algorithm includes: Taking the planning scheme as an individual, initialize multiple populations and construct an iterative community; Based on the preset fitness function, the fitness value of each individual is calculated. Based on the preset number of elites, the selection ratio is allocated according to the fitness value, and the elite individuals are determined by random proportion. Through the crossover operation, according to the preset crossover number, the planning schemes of any pair of elite individuals are partially exchanged to generate new individuals; through the mutation operation, according to the preset mutation number, any other individuals are selected to randomly explore in the solution space to generate mutant individuals, and the elite individuals, the new individuals and the mutant individuals are formed into a first type of new population; Based on the preset migration interval, after the corresponding rounds of iterations, according to the preset migration ratio, the corresponding number of individuals with the largest fitness values are selected from each population as migration individuals, and the individuals are exchanged and migrated among all populations to generate the second type of new population; according to the preset alliance size, the corresponding number of populations are merged to generate the third type of new population; The first type of new population, the second type of new population and the third type of new population together form a new community, and the iteration continues; Until the preset number of iterations is reached, the individual with the largest fitness value is selected from the final iterative community to obtain the optimal planning solution.
2. The method according to claim 1, characterized in that Based on the current node, the node degree is calculated, the return factor is determined, and the clustering coefficient of the neighborhood is calculated, the exploration factor is determined, and the weights of all the connecting edges connected to the current node are combined to determine the transition probability distribution corresponding to the neighboring node, including: The transition probability from the current node to the neighboring node is as follows: ; in, u Representation Node u , v Represents another node v , P ( u | v ) indicates a slave node v To Node u The transition probability, β represents the balance factor, deg( v ) represents a node v The node degree of N ( v ) represents a node v The set of neighboring nodes of x Represents a node in the neighborhood node set x , C u Representation Node u The clustering coefficient of the neighborhood, C x Representation Node x The clustering coefficient of the neighborhood, ω vu Representation Node v and nodes u The weight of the edge connecting them, y Represents a node in the neighborhood node set y , C y Representation Node y The clustering coefficient of the neighborhood, ω vx Representation Node v and nodes x The weight of the connecting edge.
3. The method according to claim 1, characterized in that: By maximizing the loss function and iteratively adjusting the model parameters until the preset number of iterations is reached, the optimal resource demand prediction network is generated, including: The loss function has the following formula: ; in, L represents the loss function, i represents the Gaussian component ordinal number, M represents the total number of mixed distributions, π i Indicates i The mixing coefficients of Gaussian components, p represents the variables in the input sequence, q represents the variables in the output sequence, μ i Indicates i The mean of the Gaussian components, σ i Indicates i The standard deviation of the Gaussian components.
4. A system resource requirement planning system based on artificial intelligence, used to implement the method described in any one of claims 1 to 3, characterized in that: include: The first unit is used to obtain system resource data, construct a resource attribute heterogeneous graph, generate a resource node sequence set through multiple rounds of random walks based on the resource attribute heterogeneous graph, learn the node neighborhood relationship through a training graph embedding model, construct a candidate feature matrix, and generate an optimal feature subset based on the comprehensive feature score of each candidate feature; The second unit is used to input the system resource time series data to be predicted into the resource demand prediction network, generate mixed Gaussian distribution parameters by capturing the dependency relationship of each time step, determine the conditional probability distribution, and obtain the resource demand prediction result; The third unit is used to randomly generate multiple planning schemes according to the resource demand prediction results and the system resources determined based on the optimal feature subset, construct an iterative community, input the iterative community into the resource allocation algorithm, generate the optimal solution, and determine the optimal planning scheme.
5. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 3.
6. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 3 is implemented.
