Project job resource prediction method based on LSTM-GAT

Abstract

The invention belongs to the technical field of geophysical exploration, and particularly discloses a project job resource prediction method based on LSTM-GAT. The method comprises the following steps: constructing a job resource utilization network through a threshold method, extracting a project resource utilization state by utilizing a gating loop unit network, establishing a full-connection network for each category, firstly aggregating state information in project job categories through a graph attention network, and establishing a full-connection network for each category; and aggregating the state information among the classes on the basis of the state information in the classes, and finally merging the state information in the classes and the state information among the classes to serve as final project resource utilization information. According to the method, the use efficiency of the resources can be improved, the queuing waiting time of the project is shortened, and the resources required by the project are effectively predicted.

Description

Technical Field

[0001] The present invention belongs to the field of computer application technology, and in particular relates to a project operation resource prediction method based on LSTM-GAT. Background Art

[0002] When the job resource management software collects and processes resources of nodes in the cluster and customizes job scheduling and resource management strategies, if it only considers the resource situation at the time of job submission, then the understanding of the entire system will only stay at the historical and current resource usage of the entire system, and there is no certain analysis and estimation of the resource utilization of the entire system in the future and the situation that will occur, so it is impossible to predict unexpected situations that may occur in the system. Demand forecasting for resources required for a project is an effective solution for rapid resource allocation and then resource management and intelligent scheduling. In recent years, resource forecasting algorithms and systems have become indispensable for appropriate resource allocation in clusters.

[0003] Existing prediction algorithms, such as the multi-layer perceptron model (MLP), as the simplest form of neural network, also use deep belief networks (DBN), as well as hybrid deep learning frameworks of graph convolutional networks (GCN) and long short-term memory networks (LSTM). Some recent studies have also focused on longer-term predictions, such as deep learning stacked autoencoder methods, or using LSTM neural networks to capture long-term temporal dependencies. The disadvantage of the MLP model is that it is too simple. Once the data set is not linearly separable, its application will be greatly limited. The DBN model is a probabilistic model, so its prediction accuracy is not high. The hybrid deep learning framework of graph convolutional networks (GCN) and long short-term memory networks (LSTM) has a slow training speed and requires a lot of computing resources, and sometimes overfitting occurs. The deep learning stacked autoencoder method often requires a lot of data, a long training time, and requires the adjustment of a large number of hyperparameters. The LSTM neural network has disadvantages in parallel processing and is time-consuming to calculate.

[0004] The emergence of Graph Attention Network (GAT) modifies the convolution operation in GCN to compute the hidden representation of each node in the graph by paying attention to the neighbors of the node, thereby performing node classification on the graph data structure. Therefore, GAT performs the functions of GCN while adapting itself to the most important specific features in the network. GAT also promises efficient and parallelizable operations, applied to graph nodes of different degrees, and suitable for inductive learning problems. Summary of the invention

[0005] In order to solve the above deficiencies in the known technology, the present invention aims to provide a project operation resource prediction method based on LSTM-GAT, so as to improve the efficiency of resource utilization, reduce the waiting time of project queues, and effectively predict the resources required for the project.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0007] A project operation resource prediction method based on LSTM-GAT, the method comprising the following steps performed in sequence:

[0008] P1. Calculate the standardized mutual information between project operation resource requirements, use the standardized mutual information to measure the relevance of project operations, and construct the project operation resource utilization network through the threshold method;

[0009] P2, using the gated recurrent unit network to learn the status of each project operation within the project operation resource utilization time window and determine the type of project operation;

[0010] P3, treat the same type of project work as a separate category, and build a fully connected network for each category;

[0011] P4. Use the graph attention network to learn the relationship between the same type of project tasks, input the state within the resource utilization time window as the initial state into the graph attention network, and input the edge relationship between project resource utilization into the graph attention network;

[0012] P5. After the graph attention network learns the attention weights between the resource utilization of project jobs within a category, it uses the attention weights to aggregate the information within the same category and corrects it through the activation function ReLU to obtain the in-category state of each project job after weighted correction by the graph attention network.

[0013] P6. Treat each category as a node, construct a category fully connected network, use the attention mechanism based on the feedforward neural network to learn the attention weights of the project tasks in each category, and aggregate the inter-category state of each category through the weighted intra-category states of all project tasks in each category;

[0014] P7, input the inter-category states of all categories into the graph attention network, and obtain the weighted and corrected inter-category states through mutual weighted aggregation between categories;

[0015] P8. The weighted corrected intra-category state and the weighted corrected inter-category state of project resource utilization are concatenated, and then corrected by linear transformation and activation function. The state information after linear transformation and activation function correction is used as the final project resource utilization state information. The final features generated by the project in the time window are:

[0016]

[0017] in, is the category to which the item s belongs, Wf is the learnable parameter matrix, is the weight of item s in the graph attention network, is the weight of item s in the feedforward neural network, τ i (π c ) is the category π c In the resource utilization time window to the weighted modified inter-category state, when the project belongs to multiple categories, τ i (π c ) takes the average of the inter-category states of multiple categories to which item s belongs.

[0018] As a limitation, in step P1, a numerical analysis method is used to calculate the standardized mutual information between job resource requirements, and the calculation method is performed in sequence according to the following steps:

[0019] a1) The parameter feature Si[s1,s2,…sn] of item i corresponds to the item parameter interval [minSi,maxSi]. The item parameter feature interval is divided into K sub-intervals, and the frequency f of the parameter feature Si of item i falling in the kth sub-interval is calculated. i,k , using frequency to approximate probability p i,k :

[0020]

[0021] Where d is the sample size, the entropy of Si is:

[0022]

[0023] a2) Let the parameter interval corresponding to the parameter feature Sj[s1,s2,…sn] of project j be [minSj,maxSj]. For the joint entropy, divide [minSi,maxSi]*[minSj,maxSj] into K*K subintervals, and calculate the frequency f of the joint yield (Si,Sj) of project i and project j falling in the subinterval (k,l) i,j,k,l , using frequency to approximate probability p i,j,k,l :

[0024]

[0025] The joint entropy of the joint rate of return (Si, Sj) is:

[0026]

[0027] Substitute H(Si) and H(Si,Sj) into the calculation formula of the normalized mutual information to obtain the approximate value of the normalized mutual information, which is used to represent the correlation between items.

[0028] As a second limitation, the threshold method in step P1 controls the filtering of network information by adjusting the threshold. When the normalized mutual information between two project parameters does not exceed the threshold, the corresponding edge between the two project parameter nodes is deleted.

[0029] As a third limitation, the state of each project operation in the project operation resource utilization time window in step P2 is expressed as:

[0030] h i =LSTM(v i )

[0031] v i = {v i1 ,v i2 ,...,v in},

[0032] Among them, h i is the output state of project i within the project resource utilization time window, v i Represents the input parameter status of project i within the project resource utilization time window.

[0033] As a fourth limitation, the attention weights between the intra-category project tasks in step P5 are expressed as:

[0034]

[0035] The weighted and corrected in-category status of each project job through the graph attention network is represented as:

[0036] g s =GAT(s)=ReLU(∑ s'∈Γ(s) β ss' W s' ),

[0037] where β ss' represents the attention weight of project s' to project s within the project resource utilization time window, g s represents the in-category state of item s after weighted correction by the graph attention network, GAT represents the graph attention network, Γ(s) represents all the neighboring items of item s and itself, W is a learnable shared linear transformation matrix, r is a learnable shared attention mechanism, r T is the transpose of matrix r, h is the output state within the project resource utilization time window, and project s and project s' are two adjacent projects in project i.

[0038] As a fifth limitation, the attention weight of the project work in each category in step P6 is:

[0039]

[0040] in, Represents category π c Medium Project The attention weight, is the category π c The corresponding learnable parameter matrix;

[0041] The inter-category status for each category is:

[0042]

[0043] in, For the πth c The inter-category status of each category within the resource utilization time window;

[0044]

[0045] in, For category π c All project assignments included in |π c | is the category π c The number of project jobs included in is the category π c The collection of all project jobs in the .

[0046] As a sixth limitation, in step P7, the inter-category states of all categories are input into the graph attention network, and the weighted and corrected inter-category states are obtained by mutual weighted aggregation between categories, which can be expressed by the formula:

[0047]

[0048] in,

[0049] is the class π to the class π c The attention weight, W π is a learnable shared linear transformation matrix for class π, r π is a learnable shared attention mechanism for category π, is the transpose of the learnable shared attention mechanism for category π, G π is a class fully connected network, and π' is a class different from class π.

[0050] As a seventh limitation, the project operation resources in step P1 include the number of nodes, CPU, GPU, storage, and completion time.

[0051] Due to the adoption of the above technical solution, the present invention has the following beneficial effects compared with the prior art:

[0052] (1) The method of the present invention achieves better management of resources and projects, improves the efficiency of resource utilization, reduces the waiting time for projects, and effectively predicts the resources required for projects;

[0053] (2) The method of the present invention assists in realizing intelligent task scheduling. According to the historical forms submitted by the applicants, the data is analyzed and a prediction model is established to predict the number of nodes, CPU, GPU, storage, completion time, etc. required for each task (job).

[0054] In summary, the present invention can improve the efficiency of resource utilization, reduce the waiting time for projects, and achieve effective prediction of resources required for projects. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0056] Figure 1 This is a flow chart of a project operation resource prediction model according to an embodiment of the present invention;

[0057] Figure 2 It is a diagram showing the project operation of an embodiment of the present invention;

[0058] Figure 3 This is a schematic diagram of the LSTM structure of an embodiment of the present invention;

[0059] Figure 4 Schematic diagram of the attention network structure of an embodiment of the present invention;

[0060] Figure 5 Schematic diagram of the structure of a feedforward neural network FNN according to an embodiment of the present invention. DETAILED DESCRIPTION

[0061] In order to better explain the present invention and facilitate understanding, the preferred embodiments of the present invention are described in detail below through specific implementation modes in conjunction with the accompanying drawings.

[0062] Example Project operation resource prediction method based on LSTM-GAT

[0063] This embodiment provides a project operation resource prediction method based on LSTM-GAT. Figure 1 The figure shows a flow chart of the project operation resource prediction model of this embodiment. A gated recurrent unit network is used to extract the status of project operation resource utilization. The graph attention network is used to first aggregate the status information within the project operation category. Then, based on the status information within the category, the status information between categories is aggregated. Finally, the status information within the category and the status information between categories are merged as the final project resource utilization information. Multi-task learning is performed, and a feedforward neural network is used to predict project resource utilization.

[0064] The project operation resource prediction method based on LSTM-GAT in this embodiment includes the following steps performed in sequence:

[0065] P1. Calculate the standardized mutual information between project job resource requirements, use the standardized mutual information to measure the correlation of job features, and construct a project job resource utilization network through the threshold method. Project job resources include the number of nodes, CPU, GPU, storage, and completion time.

[0066] P2. Use the gated recurrent unit network to learn the status of each project operation within the project operation resource utilization time window and determine the type of project operation. The status of each project operation within the project operation resource utilization time window can be expressed as:

[0067] h i =GRU(v i )

[0068] v i = {v i1 ,v i2 ,…,v in},

[0069] Among them, h i is the output state of project i within the project resource utilization time window, v i Represents the input parameter status of project i within the project resource utilization time window. It should be noted that the status of each project operation within the project operation resource utilization time window is determined by the historical data of project resource utilization.

[0070] The Gated Recurrent Unit network has a special mechanism to determine when the state should be updated and when the state should be reset. When the reset gate is open, the Gated Recurrent Unit contains the basic recurrent neural network; when the update gate is open, the Gated Recurrent Unit can skip subsequences;

[0071] P3, treat the same type of project work as a separate category, and build a fully connected network for each category;

[0072] P4. Use the graph attention network to learn the relationship between the same type of project tasks, input the state within the resource utilization time window as the initial state into the graph attention network, and input the edge relationship between project resource utilization into the graph attention network;

[0073] P5. After the graph attention network learns the attention weights between the resource utilization of project jobs within a category, it uses the attention weights to aggregate the information within the same category and corrects it through the activation function ReLU to obtain the in-category state of each project job after weighted correction by the graph attention network.

[0074] P6. Treat each category as a node, construct a category fully connected network, use the attention mechanism based on the feedforward neural network to learn the attention weights of the project tasks in each category, and aggregate the inter-category state of each category through the weighted intra-category states of all project tasks in each category;

[0075] Among them, the feedforward neural network is a basic deep learning model, which consists of multiple fully connected layers. Assume that the input vector is x∈R D , the output vector is y∈R C , where C is the output dimension. For a single-layer feedforward neural network, the calculation formula can be expressed as:

[0076] y=σ(w Τ σ(W X +b)+b′),

[0077] Where W∈R H×D is the weight matrix, b∈R H is the bias vector, W∈R H and b′ are the weights and biases of the output layer, σ(·) is the activation function, usually a nonlinear function such as ReLU, sigmoid, etc.

[0078] P7. Input the inter-category states of all categories into the graph attention network, and obtain the weighted and corrected inter-category states through mutual weighted aggregation between categories:

[0079] P8. The weighted corrected intra-category state and the weighted corrected inter-category state of project resource utilization are spliced, and then the linear transformation and activation function correction are performed. The state information after the linear transformation and activation function correction is used as the final project resource utilization state information. The new project information is input through the feedforward neural network to predict the resources required for the project. The final features generated by the project within the time window are:

[0080]

[0081] in, is the category to which the item s belongs, Wf is the learnable parameter matrix, h i =GRU(v i ) is the weight of item s in the graph attention network, h s' is the weight of item s in the feedforward neural network, τ i (π c ) is the category π c In the resource utilization time window to the weighted modified inter-category state, when the project belongs to multiple categories, τ i (π c ) takes the average of the inter-category states of multiple categories to which item s belongs.

[0082] In step P1 of this embodiment, a numerical analysis method is used to calculate the standardized mutual information between job resource requirements, and the calculation method is performed in sequence according to the following steps:

[0083] a1) The parameter feature Si[s1,s2,…sn] of item i corresponds to the item parameter interval [minSi,maxSi]. The item parameter feature interval is divided into K sub-intervals, and the frequency f of the parameter feature Si of item i falling in the kth sub-interval is calculated. i,k , using frequency to approximate probability p i,k :

[0084]

[0085] Where d is the sample size, the entropy of Si is:

[0086]

[0087] a2) Let the parameter interval corresponding to the parameter feature Sj[s1,s2,…sn] of project j be [minSj,maxSj]. For the joint entropy, divide [minSi,maxSi]*[minSj,maxSj] into K*K subintervals, and calculate the frequency f of the joint yield (Si,Sj) of project i and project j falling in the subinterval (k,l) i,j,k,l , using frequency to approximate probability p i,j,k,l :

[0088]

[0089] The joint entropy of the joint rate of return (Si, Sj) is:

[0090]

[0091] Substitute H(Si) and H(Si,Sj) into the calculation formula of the normalized mutual information to obtain the approximate value of the normalized mutual information, which is used to represent the correlation between items.

[0092] Normalized mutual information is a commonly used statistic in text mining and natural language processing, which is used to measure the degree of association between two random variables. The calculation formula of normalized mutual information is as follows:

[0093]

[0094] Where I(X,Y) is the mutual information of X and Y, H(X) and H(Y) are the entropies of X and Y. This formula can map the mutual information value to the interval [-1,1]. The closer to 1, the stronger the correlation between the two variables. The closer to -1, the weaker or negative correlation between the two variables. The closer to 0, the less obvious correlation there is between the two variables.

[0095] After using mutual information to calculate the correlation between project parameters, a fully connected network is actually constructed. In order to analyze the overall structure of the project parameters, the network needs to be filtered to extract redundant information from the network and filter out key information from the network. Commonly used filtering methods include the minimum spanning tree method, the plane maximum filtering method, and the threshold method. In this embodiment, the threshold method is used. The threshold method in step P1 controls the filtering of network information by adjusting the threshold. When the standardized mutual information between two project parameters does not exceed the threshold, the corresponding edge between the two project parameter nodes is deleted.

[0096] In step P2, projects with the same offset or using the same type of software are considered as a separate project category. Generally, projects in the same category are more correlated with each other, so the mutual influence between projects is first learned through projects within the category. In this embodiment, it is assumed that the project nodes in each category are fully connected. c Represents the cth category and establishes a fully connected graph for each category

[0097] The attention weights between intra-category item assignments in step P5 are expressed as:

[0098]

[0099] The weighted and corrected in-category status of each project job through the graph attention network is represented as:

[0100] g s =GAT(s)=ReLU(∑ s'∈Γ(s) β ss' W s' ),

[0101] where β ss' represents the attention weight of project s' to project s within the project resource utilization time window, g s represents the in-category state of item s after weighted correction by the graph attention network, GAT represents the graph attention network, Γ(s) represents all the neighboring items of item s and itself, W is a learnable shared linear transformation matrix, r is a learnable shared attention mechanism, r T is the transpose of matrix r, h is the output state within the project resource utilization time window, and project s and project s' are two adjacent projects in project i.

[0102] The attention weights of the project tasks within each category in step P6 are:

[0103]

[0104] in, Represents category π c Medium Project The attention weight, is the category π c The corresponding learnable parameter matrix;

[0105] The inter-category status for each category is:

[0106]

[0107] in, For the πth c The inter-category status of each category within the resource utilization time window;

[0108]

[0109] in, For category π c All project assignments included in |π c | is the category π c The number of project jobs included in is the category π c The collection of all project jobs in the .

[0110] To learn the attention weights, in the fully connected network of each category, a feedforward neural network-based attention mechanism is used to learn the attention weights of the project tasks within each category. This can be done by inputting the state of the project tasks into the attention network to generate an attention weight vector. The attention weight vector can be weighted according to the importance of the project tasks to highlight the impact of key project tasks.

[0111] In step P7, the inter-category states of all categories are input into the graph attention network, and the weighted and corrected inter-category states are obtained by mutual weighted aggregation between categories, which can be expressed as:

[0112]

[0113] in,

[0114] is the class π to the class π c The attention weight, W π is a learnable shared linear transformation matrix for class π, r πis a learnable shared attention mechanism for category π, is the transpose of the learnable shared attention mechanism for category π, G π is a class fully connected network, and π' is a class different from class π.

[0115] Figures 2 to 5 The model of the project resource forecasting process is shown in FIG. Figures 2 to 5 model to obtain the resource space required for the job. Figure 2 Shown is a project activity representation diagram showing activity data related to resource space allocation.

[0116] Figure 3 Figure 1 is a schematic diagram of the structure of LSAM. LSTM is a recursive neural network that is often used to process and predict time series data. Unlike traditional recurrent neural networks, LSTM has the ability to remember or forget certain information for a long time, and is particularly suitable for processing sequence data with long-term dependencies, such as speech recognition or natural language processing. LSTM units are usually composed of three gating units that control the inflow and outflow of information, including the input gate, the forget gate, and the output gate. These gating units allow LSTM to selectively store and retrieve relevant information from previous time steps, making them an effective solution for many sequence modeling tasks.

[0117] Figure 3 In the , SM is soft maximum, a function that can compress any real vector into a probability distribution of several possible discrete results. It is often used as the activation function of the output layer in classification tasks; FC is a fully connected layer, a neural network layer in which every neuron in one layer is connected to every neuron in the next layer. This layer type is also called a dense or linear layer; L is an LSTM layer, a single LSTM unit with an input, output, and forget gate; C is a convolutional layer, a neural network layer that applies filters to local areas of the input data to extract features and learn representations. These layers are often used in computer vision tasks.

[0118] Figure 4 The following is a diagram of the graph attention network structure. The graph attention network is a deep learning model for graph structured data. Unlike traditional models such as convolutional neural networks (CNN) and recurrent neural networks (RNN) that target sequence or grid data, it can effectively process graph data with arbitrary topological structures. Graph attention networks usually consist of two parts: node representation and edge representation. The node representation is used to learn node features, and the edge representation is used to establish relationships between nodes. In the graph attention network, both nodes and edges are assigned weights, namely attention coefficients. These weights can be adjusted dynamically so that the network can focus on the most relevant nodes and edges. Among them, represents the input features, represents the attention weight of the node to the node, It is the final feature after processing by this node. Figure 5 The diagram is a structural diagram of a feedforward neural network FNN. FNN is a basic neural network model consisting of an input layer, several hidden layers, and an output layer. In FNN, information flows only from the input layer to the output layer, and there is no feedback connection.

Claims

1. A project operation resource prediction method based on LSTM-GAT, characterized in that: The method comprises the following steps performed in sequence: P1. Calculate the standardized mutual information between project operation resource requirements, use the standardized mutual information to measure the relevance of project operations, and construct the project operation resource utilization network through the threshold method; P2, using the gated recurrent unit network to learn the status of each project operation within the project operation resource utilization time window and determine the type of project operation; P3, treat the same type of project work as a separate category, and build a fully connected network for each category; P4. Use the graph attention network to learn the relationship between the same type of project tasks, input the state within the resource utilization time window as the initial state into the graph attention network, and input the edge relationship between project resource utilization into the graph attention network; P5. After the graph attention network learns the attention weights between the resource utilization of project jobs within a category, it uses the attention weights to aggregate the information within the same category and corrects it through the activation function ReLU to obtain the in-category state of each project job after weighted correction by the graph attention network. P6. Treat each category as a node, construct a category fully connected network, use the attention mechanism based on the feedforward neural network to learn the attention weights of the project tasks in each category, and aggregate the inter-category state of each category through the weighted intra-category states of all project tasks in each category; P7, input the inter-category states of all categories into the graph attention network, and obtain the weighted and corrected inter-category states through mutual weighted aggregation between categories; P8. The weighted corrected intra-category state and the weighted corrected inter-category state of project resource utilization are concatenated, and then corrected by linear transformation and activation function. The state information after linear transformation and activation function correction is used as the final project resource utilization state information. The final features generated by the project in the time window are: in, is the category to which the item s belongs, Wf is the learnable parameter matrix, is the weight of item s in the graph attention network, is the weight of item s in the feedforward neural network, τ i (π c ) is the category π c In the resource utilization time window to the weighted modified inter-category state, when the project belongs to multiple categories, τ i (π c ) takes the average of the inter-category states of multiple categories to which item s belongs.

2. The project operation resource prediction method based on LSTM-GAT according to claim 1 is characterized in that: In step P1, a numerical analysis method is used to calculate the standardized mutual information between the job resource requirements, and the calculation method is performed in the following steps in sequence: a1) The parameter feature Si[s1,s2,…sn] of item i corresponds to the item parameter interval [minSi,maxSi]. The item parameter feature interval is divided into K sub-intervals, and the frequency f of the parameter feature Si of item i falling in the kth sub-interval is calculated. i,k , using frequency to approximate probability p i,k : Where d is the sample size, the entropy of Si is: a2) Let the parameter interval corresponding to the parameter feature Sj[s1,s2,...sn] of project j be [minSj,maxSj]. For the joint entropy, divide [minSi,maxSi]*[minSj,maxSj] into K*K subintervals, and calculate the frequency f of the joint yield (Si,Sj) of project i and project j falling in the subinterval (k,l) i,j,k,l , using frequency to approximate probability p i,j,k,l : The joint entropy of the joint rate of return (Si, Sj) is: Substitute H(Si) and H(Si,Sj) into the calculation formula of the normalized mutual information to obtain the approximate value of the normalized mutual information, which is used to represent the correlation between items.

3. The project operation resource prediction method based on LSTM-GAT according to claim 2 is characterized in that: The threshold method in step P1 controls the filtering of network information by adjusting the threshold. When the normalized mutual information between two project parameters does not exceed the threshold, the corresponding edge between the two project parameter nodes is deleted.

4. The project operation resource prediction method based on LSTM-GAT according to claim 3 is characterized in that: The status of each project operation in the project operation resource utilization time window in step P2 is expressed as: h i =GRU(v i ) v i ={v i1 ,v i2 ,...,v in }, Among them, h i is the output state of project i within the project resource utilization time window, v i Represents the input parameter status of project i within the project resource utilization time window.

5. The project operation resource prediction method based on LSTM-GAT according to claim 4 is characterized in that: The attention weights between the intra-category tasks in step P5 are expressed as: The weighted and corrected in-category status of each project job through the graph attention network is represented as: g s =GAT(s)=ReLU(∑ s'∈Γ(s) β ss' Wh s' ), where β ss' represents the attention weight of project s' to project s within the project resource utilization time window, g s represents the in-category state of item s after weighted correction by the graph attention network, GAT represents the graph attention network, Γ(s) represents all the neighboring items of item s and itself, W is a learnable shared linear transformation matrix, r is a learnable shared attention mechanism, r T is the transpose of matrix r, h is the output state within the project resource utilization time window, and project s and project s' are two adjacent projects in project i.

6. The project operation resource prediction method based on LSTM-GAT according to claim 5 is characterized in that: The attention weight of the project work in each category in step P6 is: in, Represents category π c Medium Project The attention weight, is the category π c The corresponding learnable parameter matrix; The inter-category status for each category is: in, For the πth c The inter-category status of each category within the resource utilization time window; in, For category π c All project assignments included in |π c | is the category π c The number of project jobs included in is the category π c The collection of all project jobs in the .

7. The project operation resource prediction method based on LSTM-GAT according to claim 6 is characterized in that: In step P7, the inter-category states of all categories are input into the graph attention network, and the weighted and corrected inter-category states are obtained by mutual weighted aggregation between categories, which can be expressed as: in, is the class π to the class π c The attention weight, W π is a learnable shared linear transformation matrix for class π, r π is a learnable shared attention mechanism for category π, is the transpose of the learnable shared attention mechanism for category π, G π is a class fully connected network, and π' is a class different from class π.

8. The project operation resource prediction method based on LSTM-GAT according to claim 1 is characterized in that: The project operation resources in step P1 include the number of nodes, CPU, GPU, storage, and completion time.

Images