A time sequence feature enhanced oil field catastrophe scene emergency resource scheduling intelligent agent decision optimization method and system

Abstract

The application relates to a time sequence feature enhanced oil field catastrophe scene emergency resource scheduling intelligent agent decision optimization method and system, and relates to the field of catastrophe scene emergency resource scheduling intelligent agent decision optimization. The method solves the problems of insufficient time sequence feature mining and weak representation in the existing oil field catastrophe scene emergency resource scheduling intelligent agent decision optimization field. The method applies TCN causal convolution and self-attention mechanism to oil field catastrophe time sequence feature extraction, constructs a standardized oil field catastrophe synthetic data generation model, simulates the time sequence evolution characteristics of two typical catastrophes, namely fire and leakage, constructs a double-dimensional reward function fusing resource utilization rate and response time, makes the training target of the reinforcement learning intelligent agent highly match the actual demand of the oil field catastrophe emergency, deeply fuses the reinforcement learning intelligent agent and integer programming, and realizes adaptive decision of the catastrophe dynamic time sequence scene based on the enhanced time sequence feature, and outputs a dynamic resource scheduling scheme in combination with the integer programming.

Description

Technical Field

[0001] This invention relates to the field of intelligent agent decision optimization for emergency resource scheduling in oilfield disaster scenarios, specifically to a method and system for intelligent agent decision optimization for emergency resource scheduling in oilfield disaster scenarios with enhanced temporal characteristics. Background Technology

[0002] Emergency resource allocation in oilfield disasters is a core component of emergency management for safe production in oilfields, directly determining the efficiency of disaster response and the effectiveness of loss control. However, existing oilfield disaster resource allocation technologies still have many core shortcomings and are difficult to match the dynamic allocation needs under the complex temporal evolution scenarios of oilfield disasters. Insufficient mining of temporal features and weak representation ability: Traditional resource scheduling only uses original static features such as the speed of disaster spread and the scope of impact, without mining the temporal correlation and evolution pattern of disaster data. The feature dimension is low and the representation ability is weak, which cannot accurately reflect the differences in resource demand at different stages of disaster. Limited data support and model training: Real disaster data from oilfields is scarce due to the special nature of the scenarios and incomplete records. There is also no standardized mechanism for generating synthetic disaster data. Decision models that rely solely on real data for training have poor generalization ability and are difficult to adapt to different types and levels of oilfield disaster scenarios. The decision-making methods are too simplistic and the optimization results are poor: existing technologies either use static scheduling methods based on pure integer programming, which cannot adapt to the dynamic temporal evolution of disasters; or they use a single reinforcement learning agent for decision-making, which lacks global optimal constraints at the mathematical level, and cannot simultaneously achieve dynamic adaptability and global optimization of resource scheduling. The poor design of the reward mechanism leads to poor training effect of the agent: the reward function of some reinforcement learning scheduling schemes only considers the resource utilization rate or response time, and fails to build a multi-dimensional scientific reward mechanism, which causes the decision-making strategy of the agent training to be out of touch with the actual needs of oilfield disaster emergency response.

[0003] In summary, existing intelligent agent decision-making optimization methods for emergency resource scheduling in oilfield disaster scenarios suffer from several problems, including insufficient mining of temporal features, weak representation capabilities, limited data support, restricted model training, simplistic decision-making methods, poor optimization results, unreasonable reward mechanism design, and poor agent training performance. Summary of the Invention

[0004] This invention addresses the shortcomings of existing intelligent agent decision-making optimization methods for emergency resource scheduling in oilfield disaster scenarios, including insufficient temporal feature mining, weak representation capabilities, limited data support, restricted model training, simplistic decision-making methods, poor optimization results, and unreasonable reward mechanism design, leading to poor agent training performance. To resolve these issues, this invention proposes a temporal feature-enhanced intelligent agent decision-making optimization method and system for oilfield disaster scenarios. The invention achieves these results through the following technical solutions: Option 1: This invention proposes a decision optimization method for an intelligent agent in emergency resource scheduling for oilfield disaster scenarios, enhanced with time-series features. The method includes the following steps: Step 1: Configure data parameters in the visualization interface, construct dual-source data for synthetic data generation and real data processing based on the selected data parameters, and preprocess the dual-source data; Step 2: Extract the disaster spread speed, impact range, personnel demand, equipment demand, disaster level, and time step characteristics from the dual-source data to form an original time series feature matrix. Normalize the original time series feature matrix and convert the normalized feature matrix into a tensor format that the model can input, and output the normalized time series feature tensor. Step 3: Input the normalized temporal feature tensor into the TCN causal convolution and self-attention mechanism, extract the local correlation features of the oilfield disaster time series data through the TCN causal convolution layer in sequence, expand the feature range with an exponential expansion rate, and enhance the weight of the extracted local features through the self-attention mechanism to output the enhanced high-dimensional temporal features. Step 4: Construct a reinforcement learning environment for oilfield disaster resource scheduling based on the enhanced high-dimensional temporal features, define the state space, action space and two-dimensional reward function, and train the agent model. Step 5: Extract the real-time state of the catastrophe from the enhanced high-dimensional temporal features, input it into the agent model trained in Step 4, and output the optimal scheduling action. Step 6: Perform integer programming resource scheduling on the optimal scheduling action, construct an integer programming model based on the real-time state of the disaster, obtain the globally optimal resource scheduling scheme, feed the scheduling result back to the reinforcement learning environment, update the agent model, output the dynamic resource scheduling scheme, and display the scheduling result through a visualization interface.

[0005] Furthermore, a preferred embodiment is provided, wherein the method for constructing dual-source data of synthetic data generation and real data processing in step 1 is as follows: the synthetic data is used to generate time-series data with labeling information such as disaster level and resource demand by simulating the time-series evolution of typical disasters such as oilfield fires and oil and gas leaks; the real data is used to perform missing value filling and outlier filtering standardization processing on the missing values ​​and outliers existing in the oilfield field data in sequence.

[0006] Furthermore, a preferred embodiment is provided, wherein the method for extracting local correlation features of oilfield disaster time series data through a TCN causal convolutional layer in step 3 is as follows: Let the input timing characteristics be The causal convolution output is Y. but: , The kernel size k ranges from k∈[3,7], and the exponential expansion rate d ranges from d∈[1,8]. Combined with residual connections: , Output for residual adaptation layer.

[0007] Furthermore, a preferred embodiment is provided, wherein the method for weight enhancement of the extracted local features using a self-attention mechanism in step 3 is as follows: Let the features extracted by TCN be If the self-attention output is H′, then: , in, is the matrix dimension.

[0008] Furthermore, a preferred implementation method is provided, wherein the method for defining the state space, action space, and two-dimensional reward function in step 4 and training the agent model is as follows:

[0009] in This represents the average utilization rate of personnel and equipment resources. ; The average response time for personnel and equipment resources is used, and the DQN and PPO algorithms are employed to train the agent model.

[0010] Furthermore, a preferred implementation is provided, wherein the method for extracting the real-time catastrophic state from the enhanced high-dimensional temporal features in step 5, inputting it into the agent model trained in step 4, and outputting the optimal scheduling action is as follows: Construct the objective function:

[0011] in, For personnel response time, For personnel utilization rate, For equipment response time, equipment utilization rate To optimize variables; Set constraints: Total resource constraint:

[0012] Based on the constructed objective function and constraints, the high-dimensional interior point method is used to solve the integer programming model, and the optimal scheduling personnel, equipment allocation, utilization rate and response time are output.

[0013] Furthermore, a preferred embodiment is provided, wherein the method for outputting the dynamic resource scheduling scheme in step 6 is as follows: Step 6.1: The agent model is based on the enhanced temporal features. Output the optimal scheduling action ; Step 6.2, Substituting proportional constraints and considering the disaster state at time step t, calculate the integer programming model and output the resource scheduling scheme at time step t. ; Step 6.3: Develop a resource scheduling plan. The utilization rate and response time are substituted into the reward function to calculate the reward value and feed it back to the reinforcement learning environment to update the agent network. Step 6.4: Repeat steps 6.1-6.3 to achieve dynamic resource scheduling under the evolution of disaster time sequence.

[0014] Option 2: A decision-making optimization system for emergency resource scheduling in oilfield disaster scenarios with enhanced temporal characteristics, the system comprising: The dual-source data processing module is used to configure data parameters in a visual interface, construct synthetic data generation and real data processing dual-source data based on the selected data parameters, and preprocess the dual-source data. The time series feature enhancement module is used to extract the disaster spread speed, impact range, personnel demand, equipment demand, disaster level, and time step characteristics from dual-source data to form an original time series feature matrix. The module then normalizes the original time series feature matrix, converts the normalized feature matrix into a tensor format that the model can input, and outputs a normalized time series feature tensor. The reinforcement learning agent training module is used to input the normalized temporal feature tensor into the TCN causal convolution and self-attention mechanism. The local correlation features of the oilfield disaster time series data are extracted sequentially through the TCN causal convolution layer. The feature range is expanded by an exponential expansion rate, and the extracted local features are weighted and enhanced through the self-attention mechanism to output the enhanced high-dimensional temporal features. The integer programming resource scheduling module is used to construct a reinforcement learning environment for oilfield disaster resource scheduling based on enhanced high-dimensional temporal features, define the state space, action space and two-dimensional reward function and train the agent model; The dynamic decision optimization fusion module is used to extract the real-time state of the disaster from the enhanced high-dimensional time series features, input it into the intelligent agent model trained in the integer planning resource scheduling module, and output the optimal scheduling action. The integrated visualization monitoring module is used to perform integer planning resource scheduling for the optimal scheduling action. It combines the real-time disaster status to build an integer planning model, obtains the globally optimal resource scheduling scheme, feeds the scheduling results back to the reinforcement learning environment, updates the agent model, outputs a dynamic resource scheduling scheme, and displays the scheduling results through a visualization interface.

[0015] Option 3: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described in Option 1.

[0016] Option 4: A computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the method described in Option 1.

[0017] The advantages of this invention are: The present invention provides a method and system for optimizing decision-making of an intelligent agent for emergency resource scheduling in oilfield disaster scenarios with enhanced temporal features. For the first time, it integrates TCN causal convolution and self-attention mechanism to extract temporal features of oilfield disasters. TCN ensures the unidirectionality of temporal features and expands the receptive field, while the self-attention mechanism enhances the feature weights of key disaster stages, thereby improving feature dimensionality and representation capabilities, and accurately mining the evolutionary patterns of disaster temporal events.

[0018] The present invention describes a time-series-enhanced intelligent agent decision-making optimization method and system for emergency resource scheduling in oilfield disaster scenarios. It constructs a standardized synthetic data generation model for oilfield disasters, simulating the time-series evolution characteristics of two typical disasters: fire and leakage. At the same time, it realizes automated cleaning and standardization of real data, seamless switching between two data sources, effectively solves the problem of scarcity of real data for oilfield disasters, and significantly improves the generalization ability of the decision-making model.

[0019] The present invention provides a time-series-enhanced method and system for optimizing decision-making in emergency resource scheduling for oilfield disaster scenarios. This method breaks through the traditional single-dimensional reward design by constructing a dual-dimensional reward function that integrates resource utilization and response time. This makes the training objectives of the reinforcement learning agent highly matched with the actual needs of oilfield disaster emergency response, thereby improving the rationality of the agent's decision-making strategy and the training convergence speed.

[0020] The present invention provides a time-series feature-enhanced emergency resource scheduling intelligent agent decision optimization method and system for oilfield disaster scenarios. For the first time, it deeply integrates reinforcement learning intelligent agents with integer programming. The intelligent agent realizes adaptive decision-making in dynamic time-series disaster scenarios based on the enhanced time-series features, while integer programming realizes the mathematical global optimum of resource scheduling based on the intelligent agent's actions. The two complement each other's advantages, taking into account both the dynamic adaptability and global optimization of scheduling.

[0021] The present invention discloses an emergency resource scheduling intelligent agent decision optimization method and system for oilfield disaster scenarios with enhanced time-series characteristics. Based on the enhanced disaster time-series characteristics, the system adjusts the decision-making actions of the reinforcement learning intelligent agent in real time with the disaster time step, and outputs a dynamic resource scheduling scheme by combining integer programming. This enables resource demand adaptation at different stages such as the initial stage, development stage, and peak stage of the disaster, and solves the response lag problem of traditional fixed-ratio scheduling.

[0022] This invention is also applicable to application fields such as oilfield fires and oil and gas leaks, two typical disaster scenarios. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the decision optimization method for emergency resource scheduling in oilfield disaster scenarios with enhanced temporal characteristics, as described in Implementation Method 1.

[0024] Figure 2 This is an overall architecture diagram of an intelligent agent decision optimization system for emergency resource scheduling in oilfield disaster scenarios with enhanced temporal characteristics, as described in Implementation Method 2. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0026] Implementation Method 1, see [link] Figure 1 This embodiment describes a decision optimization method for emergency resource scheduling in oilfield disaster scenarios, which enhances time-series features. The method executes and shares data, forming a complete closed loop of disaster data processing, feature enhancement, decision optimization, and scheduling display. Specifically, it includes the following steps: S1: System Initialization and Parameter Configuration Start the visualization system to complete the initialization of the device and each functional module, and adapt to the CPU / GPU hardware environment; Configure core parameters in the visualization interface: time sequence length, original feature dimension, augmented feature dimension, reinforcement learning algorithm type (DQN / PPO), number of training steps, and integer programming constraints. Initialize the temporal feature enhancement model and the reinforcement learning agent network to prepare for subsequent processes.

[0027] S2: Preparation and Loading of Dual-Source Data Select the data source through the visual interface: synthetic data or real data; If synthetic data is selected: set the disaster type (fire / leak) and time step, the system will generate batch synthetic time-series data of oilfield disasters with complete annotations and output the synthetic dataset; If real data is selected: Select the path to the real data file, and the system will automatically load and complete the standardization processing of missing value filling and outlier filtering, and output standardized real data; The processed dataset is displayed in a visualization interface, supporting data preview and export.

[0028] S3: Time Series Data Preprocessing and Normalization Six core features—disaster spread speed, impact range, personnel demand, equipment demand, disaster level, and time step—are extracted from dual-source data to form an original time-series feature matrix. The original feature matrix is ​​normalized using Z-score to eliminate the influence of dimensions. The formula is: Where μ is the characteristic mean, σ is the characteristic standard deviation, and ε = 1e 8. Avoid denominators of 0; The normalized feature matrix is ​​converted into a tensor format that the model can input, and the normalized temporal feature tensor is output.

[0029] S4: Temporal Feature Enhancement Processing The normalized temporal feature tensor is input into the temporal feature enhancement module, and then passed through the TCN causal convolutional layer, self-attention layer, and fully connected layer in sequence. The TCN layer extracts temporal local correlation features, the self-attention layer enhances the weights of key temporal features, and the fully connected layer improves the feature dimension. Output enhanced high-dimensional time-series features, display the feature dimensions and feature values ​​at key time steps in a visualization interface, and support feature result export.

[0030] S5: Environment Construction and Training for Reinforcement Learning Agents A reinforcement learning environment for oilfield disaster resource scheduling is constructed based on enhanced post-temporal features, defining the state space, action space, and a two-dimensional reward function. Select the reinforcement learning algorithm (DQN / PPO), set the number of training steps, start the agent training, and display the training convergence curve in real time on the visualization interface; After training is complete, the agent model is saved, supporting local loading and secondary training of the model.

[0031] S6: Reinforcement Learning + Integer Programming Dynamic Decision Optimization Extract the real-time catastrophic state (catastrophic level and resource requirements at the current time step) from the enhanced time-series features, input it into the trained agent model, and output the optimal scheduling action; The optimal scheduling action is substituted into the integer programming resource scheduling module, and an integer programming model is constructed in combination with the real-time disaster status to obtain the globally optimal resource scheduling scheme (personnel / equipment allocation, utilization rate, response time). The scheduling results are fed back to the reinforcement learning environment to update the agent network and provide an optimization basis for the decision-making in the next time step. Output a dynamic resource scheduling scheme and display the scheduling results in a visual interface.

[0032] S7: Scheduling Results Push and Visualization Based on the scheduling results, differentiated information is displayed on the visualization interface. Bar charts show the utilization rate of personnel / equipment and response time, and line charts show the changes in resource allocation over time. The system automatically records all information for this scheduling, including scheduling time, disaster type, disaster level, time step, resource allocation, utilization, response time, and optimal action. The scheduling logs are stored in a local database, supporting subsequent querying, exporting, and tracing. Operations and maintenance personnel can perform operations such as parameter modification, model retraining, and scheduling result comparison in the interface, realizing full-process visual interaction.

[0033] Implementation Method 2, see below Figure 2 This embodiment describes a decision optimization system for emergency resource scheduling in oilfield disaster scenarios with enhanced temporal features. The system is based on temporal feature enhancement, supported by deep learning (TCN + self-attention), reinforcement learning (DQN / PPO), and integer programming. It utilizes full-process visualization as its operational platform and constructs a six-layer closed-loop architecture. Data exchange and result feedback are achieved between each layer. The specific architecture layers are as follows: Parameter configuration layer: Enables flexible setting and updating of parameters such as time series length, feature dimension, reinforcement learning training steps, and integer programming constraints, and outputs personalized disaster scheduling parameters; Dual-source data processing layer: Enables adaptive generation of synthetic time-series data of oilfield disasters and standardized loading and cleaning of real disaster data, providing data support for time-series feature enhancement and model training; Temporal Feature Enhancement Layer: Based on TCN causal convolution and self-attention mechanism, feature extraction and dimensionality enhancement are performed on the preprocessed catastrophic time series data to output enhanced temporal features with high dimensionality and high representation ability; Agent Decision Optimization Layer: Construct a reinforcement learning environment for oilfield disaster resource scheduling, train DQN / PPO agents, integrate integer programming to achieve dual-drive decision optimization for resource scheduling, and output the optimal resource scheduling scheme; Dispatch and execution layer: Based on the decision optimization results, realize the dynamic allocation and execution of resources such as personnel and equipment for oilfield disaster emergency response, and output dispatch and execution records.

[0034] I. Dual-Source Data Processing Module: For two typical disaster scenarios—oilfield fires and oil and gas leaks—a standardized dual-source data supply mechanism is constructed, consisting of synthetic data generation and real data processing. Synthetic data simulates the temporal evolution of these two typical disasters, generating time-series data with annotations such as disaster level and resource requirements. Real data supports loading and standardized cleaning of mainstream formats such as CSV, enabling seamless switching between the two data sources. This provides sufficient data for model training and feature enhancement, addressing the scarcity of real data for oilfield disasters. Synthetic data can simulate the temporal evolution of different disaster types and levels in batches. Real data features automated outlier filtering and missing value imputation, improving data processing efficiency by ≥50%.

[0035] Synthetic Data Generation: The pre-defined disaster types are two typical scenarios: oilfield fire and oil and gas leak. Based on the engineering patterns of disaster evolution in oilfields, core time-series features such as disaster spread speed, impact range, and resource demand are generated. The generation formulas and parameter definitions for each feature are as follows: Speed ​​of fire spread: t is the time step of the disaster (hours), and ξ is random noise (following a normal distribution N(0,0.02)) used to simulate the random disturbance of the spread speed caused by the fluctuation of wind speed and oil and gas concentration at the fire scene; Speed ​​of spread of the leak disaster: ξ is random noise (following a normal distribution N(0,0.01)) used to simulate random disturbances in leakage diffusion rate caused by changes in pipeline pressure and formation porosity; Scope of disaster impact: , The cumulative sum of the spread rate is ξ, and ξ is random noise (following a normal distribution N(0,0.1)) used to simulate the random disturbance of the influence range by terrain and on-site obstacles; Human resource requirements: In the formula, R is the disaster impact range, ξ is random noise that follows a normal distribution, and the calculation results are clipped to the upper and lower limits of the [5,100] interval to ensure the engineering rationality of the personnel demand values; Equipment resource requirements: In the formula, R is the range of disaster impact, ξ is random noise that follows a normal distribution, and the calculation results are clipped to the upper and lower limits of the [2,50] interval to ensure the engineering rationality of the equipment requirement values.

[0036] 1. Real Data Processing: Supports loading real disaster time-series data and field monitoring data from oilfields in CSV format. Addressing the issues of missing and outlier values ​​in oilfield field data, it sequentially performs missing value imputation and outlier filtering standardization processing. The specific steps are as follows: 1.1 Forward Imputation of Missing Values: To address missing values ​​in time-series data due to monitoring interruptions or acquisition failures, a forward imputation method is used. The nearest preceding valid monitoring value at the missing location is used as the imputation value. The formula is as follows: ,in The filling result for missing values ​​at time step t. The value is the nearest non-missing valid value before time step t; for missing values ​​at the beginning of the sequence, the first valid value in the sequence is used to fill the missing value, ensuring the integrity and continuity of the time series.

[0037] 1.2 3σ Principle Outlier Filtering: For outliers in the data caused by sensor malfunctions or signal interference, the 3σ criterion is used for identification and filtering. The specific steps are as follows: (1) For a single time-series feature sequence, calculate the mean μ and standard deviation σ of the full data of the sequence; (2) Determine the normal range of values ​​as This interval can cover 99.73% of the normal distribution range of time series data; (3) For abnormal values ​​that exceed the normal range, the average of the two adjacent valid values ​​before and after the time step is used to replace them to avoid the abnormal values ​​from interfering with subsequent feature extraction and model training, and to ensure the validity of the data.

[0038] II. Temporal Feature Enhancement Module: Targeting two typical disaster scenarios—oilfield fires and oil and gas leaks—this module integrates TCN causal convolution and self-attention mechanisms. First, TCN extracts local correlation features from the oilfield disaster temporal data, expanding the receptive field with an exponential expansion rate. Then, the self-attention mechanism enhances the weights of the extracted local features, highlighting key stages such as the disaster peak and abrupt change points. Finally, a fully connected layer increases the feature dimensionality, outputting enhanced high-dimensional temporal features. Core improvements: Addressing the issue of traditional feature extraction not considering the temporal correlation of disasters, TCN causal convolution ensures the unidirectionality of temporal features (preventing the leakage of future information), and the self-attention mechanism enhances the weights of key features, resulting in a feature representation capability improvement of ≥60%.

[0039] Among them, TCN causal convolution calculation: Let the input temporal features be... If the causal convolution output is Y, then: The kernel size k ranges from k∈[3,7], and the exponential expansion rate d ranges from d∈[1,8]. d increases exponentially with the number of layers in the TCN network. Combined with residual connections: , The output of the residual adaptation layer ensures feature dimension matching.

[0040] Parameter selection and performance impact explanation: Within the range of convolution kernel k∈[3,7], as the value of k increases, the receptive field coverage of local temporal features improves, and the feature extraction accuracy of catastrophic temporal mutation points can be improved by up to 12%, but the computational cost of single-step convolution increases accordingly; when k=3, both computational efficiency and basic feature extraction capability are considered, and when k=7, it is suitable for fine feature extraction scenarios of long temporal sequences. Within the expansion rate d∈[1,8], the receptive field is expanded exponentially through exponential expansion, which can cover longer catastrophe time series without increasing network depth and avoid the gradient vanishing problem. When d=1, it is suitable for extracting short time series features with high frequency changes in the early stage of a catastrophe. When d=8, it can cover long time series correlation feature mining for the entire catastrophe cycle. The maximum receptive field can reach 64 time steps, which meets the needs of capturing the time series evolution pattern of oilfield fire and oil and gas leak catastrophe within 72 hours.

[0041] Self-attention feature enhancement: Let the features extracted by TCN be... If the self-attention output is H′, then: , Where Q / K / V is the query / key / value matrix of features. Using a matrix dimension, key temporal features are highlighted through self-attention weights. Feature Dimension Enhancement: The self-attention output features are mapped to a preset enhancement dimension through a fully connected layer, outputting the enhanced features. .

[0042] III. The reinforcement learning agent training module, targeting two typical disaster scenarios—oilfield fires and oil and gas leaks—constructs a dedicated reinforcement learning environment for oilfield disaster resource scheduling. Enhanced temporal features are used as the environment state, and resource scheduling strategies as the action space. A dual-dimensional reward mechanism based on resource utilization and response time is designed. The DQN / PPO algorithm is used to train the agent, enabling it to learn the optimal scheduling actions adapted to the evolution of the disaster timeline, providing action input for subsequent decision optimization. The scientifically designed multi-dimensional reward mechanism replaces the traditional single-dimensional reward design, ensuring a high degree of match between the agent's decision-making strategy and the actual needs of oilfield disaster emergency response. The convergence speed of agent training is improved by ≥40%. Among them, the state and action space: the state space is the enhanced temporal feature S=F, and the action space is the discrete resource scheduling action A∈[0,9], corresponding to 10 different resource allocation ratios; A two-dimensional reward mechanism: The reward function comprehensively considers resource utilization and resource response time. The higher the utilization and the shorter the response time, the higher the reward value. The core formula is:

[0043] in This represents the average utilization rate of personnel and equipment resources. ; This represents the average response time (in hours) for personnel and equipment resources.

[0044] Agent training: The agent was trained using the DQN / PPO algorithms respectively. The core hyperparameters of the algorithms were set to an adjustable range to adapt to different disaster scenarios and training environments. The specific parameter ranges are as follows: Network learning rate α: The value range is α∈[0.0001,0.01], and the default setting is 0.001. A small learning rate can improve training stability and avoid network parameter oscillations, while a large learning rate can speed up training convergence. Discount factor γ: The value range is γ∈[0.9,0.99], and the default setting is 0.95. It is used to balance the agent's focus on immediate rewards and long-term benefits, and adapt to the multi-step decision-making needs under the long-term evolution of oilfield disasters. DQN algorithm experience replay pool capacity: value range is [10000, 100000], target network update step size value range is [10, 100]; PPO algorithm clipping threshold The range of values ​​is ∈[0.1,0.3], with a default value of 0.2, to ensure the stability of policy updates and avoid large-scale policy changes. During training, the agent network parameters are iteratively updated through experience replay and policy gradient optimization. The decision-making policy is continuously optimized in the interaction with the oilfield disaster scheduling environment. After training converges, the network model is saved, and the optimal scheduling action 'a' is output. .

[0045] IV. Integer Programming Resource Scheduling Module For two typical disaster scenarios—oilfield fires and oil and gas leaks—this paper constructs an integer programming model based on the optimal actions output by a reinforcement learning agent and the real-time state of the oilfield disaster (disaster level, resource requirements). The model aims to minimize resource response time and maximize resource utilization. Constraints such as total resource amount and minimum disaster level requirements are set to obtain a mathematically optimal global resource allocation scheme. The core improvement is the deep integration of integer programming with agent actions. Resource allocation ratios are determined based on agent actions, adapting the static optimization of integer programming to the dynamic temporal evolution of disasters, thus improving the global optimality of resource scheduling by ≥50%. Objective function construction: based on personnel response time Personnel utilization rate Equipment response time Equipment utilization rate To optimize the variables, the objective function is: The coefficient settings achieve the dual objectives of "minimizing response time and maximizing utilization". Constraint settings: 1. Total resource constraints:

[0046] Minimum resource requirement constraint for disaster level: In accordance with the emergency management regulations for oilfield safety production, disaster levels are divided into Level 1 (general) to Level 5 (particularly serious), i.e., disaster level L∈[1,5]. The resource allocation must meet the minimum emergency resource requirement corresponding to the disaster level. The specific constraints are as follows: Total number of emergency personnel allocated ≥ L×10, unit: person; The total amount of emergency equipment allocated is ≥L×5, unit: units / sets; this constraint ensures that the allocation of emergency resources meets the minimum safety requirements for on-site response under different levels of disaster scenarios, and avoids disaster response failure due to insufficient resource allocation; Proportional constraint: The resource allocation ratio is determined based on the agent's action 'a'. , ( For personnel needs, (For equipment requirements). The high-dimensional interior-point method is used to solve the integer programming model, and the optimal allocation of personnel / equipment, utilization rate, and response time are output.

[0047] V. Dynamic Decision Optimization Fusion Module: As the core innovative module of this invention, it constructs a dual-driven decision optimization fusion mechanism of "reinforcement learning agent + integer programming." For two typical disaster scenarios—oilfield fires and oil and gas leaks—the optimal scheduling action output by the agent is used as input to integer programming to determine the resource allocation ratio. Based on this ratio and the real-time disaster state, integer programming solves for the globally optimal resource allocation scheme, while simultaneously feeding the scheduling results back to the reinforcement learning environment, enabling continuous optimization of the agent and forming a closed-loop decision-making process of "action output - planning solution - result feedback." This overcomes the shortcomings of traditional single decision-making methods. The reinforcement learning agent achieves adaptive decision-making in dynamic disaster scenarios, while integer programming achieves the mathematical global optimum of resource scheduling. The fusion of these two approaches balances dynamism and optimality.

[0048] The agent is based on enhanced temporal features Output the optimal scheduling action at time step t. ; The integer programming module will Substituting the proportional constraints, and combining the catastrophe state (level) at time step t... Resource requirements , Solve the integer programming model and output the resource scheduling scheme at time step t. ; scheduling scheme The utilization rate and response time are substituted into the reward function to calculate the reward value and feed it back to the reinforcement learning environment to update the agent network. As the disaster progresses, repeat the above steps to achieve dynamic resource scheduling under the evolution of the disaster timeline.

[0049] VI. Integrated Visual Monitoring Module: This module provides a visual interactive interface, integrating parameter configuration, dual-source data management, time-series feature enhancement, agent model training, decision optimization, and scheduling result display—covering the entire process. It enables one-click operation and real-time visualization of oilfield disaster resource scheduling results, while automatically recording all operation and scheduling logs, supporting local data storage and traceability. Core improvements include: resolving the fragmentation issue of traditional scheduling processes by achieving integrated functionality across the entire process; intuitively displaying scheduling results such as resource utilization and response time through charts, making operation intuitive and reducing operational costs for maintenance personnel by ≥60%; and automating scheduling log recording, resolving the issue of untraceable data and results.

[0050] 1. Interactive configuration: Supports visual modification and updating of parameters such as time series length, feature dimension, training steps, and integer programming constraints; 2. Data Management: Enables parameterized generation of synthetic data, file selection and loading of real data, and supports previewing and exporting of disaster time-series data; 3. Real-time display: Display feature enhancement results, agent training convergence curves, resource scheduling utilization and response time through bar charts, line charts, etc. 4. Log recording: Automatically records data generation / loading time, feature enhancement parameters, model training information, scheduling time, disaster level, resource allocation, utilization rate, response time, etc., and supports log querying, clearing and local saving; 5. Model Management: Integrates training, saving, and loading functions for reinforcement learning agents, allowing model training and retrieval to be completed without the need for professional coding.

[0051] Example 1: To verify the technical advantages and engineering application effects of the present invention, a simulation experimental environment was constructed for typical disaster scenarios such as oilfield fires and oil and gas leaks, and multiple sets of comparative experiments were conducted. The experimental hardware environment consisted of an Intel i7-12700H CPU, an NVIDIA RTX 3060 GPU, and 32GB of memory. The software environment consisted of Python 3.9, PyTorch 2.0, and the Gurobi integer programming solver. The experimental dataset contained 1000 sets of combined disaster time-series data and 50 sets of real historical disaster monitoring data from oilfields. The experimental results are as follows: I. Verification of Temporal Feature Representation Ability and Scheduling Optimization Effect The temporal feature enhancement module and scheduling optimization method of this invention were compared with existing mainstream LSTM temporal feature extraction and single reinforcement learning scheduling methods in a comparative experiment. The experimental results show that: The TCN+ self-attention time series feature enhancement method of the present invention improves the feature characterization ability of oilfield disaster time series evolution law by ≥60%, reduces the prediction error of disaster development trend by more than 45%, and can accurately capture the feature changes of key stages such as disaster mutation point and peak period. The reinforcement learning + integer programming dual-driven decision optimization method of the present invention improves the global optimization effect of resource scheduling scheme by ≥50% and reduces the resource gap rate of disaster response by more than 70% compared with the existing single reinforcement learning scheduling method. It can simultaneously meet the adaptability to dynamic evolution of disasters and the global optimality of resource allocation.

[0052] II. Verification of Emergency Dispatch Response Efficiency and Resource Utilization A comparative experiment was conducted between the integrated scheduling system of the present invention and existing dynamic visualization-assisted decision-making technologies. The experimental results show that: Compared with existing technical solutions, the dynamic decision optimization method of the present invention reduces the response time of disaster emergency resource scheduling by ≥30%, and can complete decision optimization and solution output within 500ms after the disaster status is updated, meeting the real-time requirements of oilfield disaster emergency response. The dual-dimensional reward mechanism and integer programming optimization method of this invention improve the overall resource utilization rate of emergency personnel and equipment by ≥20%, shorten the average rescue time for disaster response by more than 25%, and significantly reduce the production losses and safety risks caused by oilfield disasters.

[0053] In summary, the time-series feature-enhanced oilfield disaster scenario emergency resource scheduling intelligent agent decision optimization method and system described in this embodiment adopts a dual-source data processing mechanism in oilfield disaster emergency scheduling scenarios, solving the industry pain point of scarce real disaster data in oilfields. Data processing efficiency is improved by ≥50%, and model generalization ability is improved by more than 40%, making it adaptable to different types and levels of oilfield disaster scenarios. Furthermore, a dual-dimensional reward mechanism is adopted, which improves the training convergence speed of the reinforcement learning intelligent agent by ≥40%, reducing the computational power consumption and time cost of model training. The integrated visualization monitoring module realizes integrated operation of the entire process, reducing the operation cost of maintenance personnel by ≥60%, while achieving traceable management of the entire scheduling process, meeting the control requirements for safe production in oilfields.

[0054] Those skilled in the art will understand that the above description is merely a preferred embodiment of the present invention, and the features described in the various embodiments and / or claims of this disclosure can be combined or combined in various ways, even if such combinations or combinations are not explicitly described in this disclosure. This is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0055] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A decision-making optimization method for emergency resource scheduling in oilfield disaster scenarios with enhanced temporal features, characterized in that, The method includes the following steps: Step 1: Configure data parameters in the visualization interface, construct dual-source data for synthetic data generation and real data processing based on the selected data parameters, and preprocess the dual-source data; Step 2: Extract the disaster spread speed, impact range, personnel demand, equipment demand, disaster level, and time step characteristics from the dual-source data to form an original time series feature matrix. Normalize the original time series feature matrix and convert the normalized feature matrix into a tensor format that the model can input, and output the normalized time series feature tensor. Step 3: Input the normalized temporal feature tensor into the TCN causal convolution and self-attention mechanism, extract the local correlation features of the oilfield disaster time series data through the TCN causal convolution layer in sequence, expand the feature range with an exponential expansion rate, and enhance the weight of the extracted local features through the self-attention mechanism to output the enhanced high-dimensional temporal features. Step 4: Construct a reinforcement learning environment for oilfield disaster resource scheduling based on the enhanced high-dimensional temporal features, define the state space, action space and two-dimensional reward function, and train the agent model. Step 5: Extract the real-time state of the catastrophe from the enhanced high-dimensional temporal features, input it into the agent model trained in Step 4, and output the optimal scheduling action. Step 6: Perform integer programming resource scheduling on the optimal scheduling action, build an integer programming model in combination with the real-time state of the disaster, obtain the globally optimal resource scheduling scheme, feed the scheduling result back to the reinforcement learning environment, update the agent model, output the dynamic resource scheduling scheme, and display the scheduling result through a visualization interface. The method for defining the state space, action space, and two-dimensional reward function in step 4 and training the agent model is as follows: in This represents the average utilization rate of personnel and equipment resources. ; The average response time for personnel and equipment resources is used, and the DQN and PPO algorithms are employed to train the agent model. In step 5, the real-time catastrophic state is extracted from the enhanced high-dimensional temporal features and input into the agent model trained in step 4. The method for outputting the optimal scheduling action is as follows: Construct the objective function: in, For personnel response time, For personnel utilization rate, For equipment response time, equipment utilization rate To optimize variables; Set constraints: Total resource constraint: Based on the constructed objective function and constraints, the high-dimensional interior point method is used to solve the integer programming model, and the optimal scheduling personnel, equipment allocation, utilization rate and response time are output.

2. The time-series feature-enhanced intelligent agent decision-making optimization method for emergency resource scheduling in oilfield disaster scenarios according to claim 1, characterized in that, The method for constructing dual-source data of synthetic data generation and real data processing described in step 1 is as follows: The synthetic data is used to generate time-series data with disaster level and resource demand labeling information by simulating the time-series evolution of typical disasters such as oilfield fires and oil and gas leaks; the real data is used to perform missing value filling and outlier filtering standardization processing on the missing values ​​and outliers in the oilfield field data in sequence.

3. The time-series feature-enhanced intelligent agent decision-making optimization method for emergency resource scheduling in oilfield disaster scenarios according to claim 1, characterized in that, The method for extracting local correlation features of oilfield catastrophe time series data through TCN causal convolutional layers in step 3 is as follows: Let the input timing characteristics be The causal convolution output is Y. but: , The kernel size k ranges from k∈[3,7], and the exponential expansion rate d ranges from d∈[1,8]. Combined with residual connections: , Output for residual adaptation layer.

4. The time-series feature-enhanced intelligent agent decision-making optimization method for emergency resource scheduling in oilfield disaster scenarios according to claim 1, characterized in that, The method for weight enhancement of the extracted local features in step 3 using a self-attention mechanism is as follows: Let the features extracted by TCN be If the self-attention output is H′, then: ， in, is the matrix dimension.

5. The time-series feature-enhanced intelligent agent decision-making optimization method for emergency resource scheduling in oilfield disaster scenarios according to claim 1, characterized in that, The method for outputting the dynamic resource scheduling scheme in step 6 is as follows: Step 6.1: The agent model is based on the enhanced temporal features. Output the optimal scheduling action ; Step 6.2, Substituting proportional constraints and considering the disaster state at time step t, calculate the integer programming model and output the resource scheduling scheme at time step t. ; Step 6.3: Develop a resource scheduling plan. The utilization rate and response time are substituted into the reward function to calculate the reward value and feed it back to the reinforcement learning environment to update the agent network. Step 6.4: Repeat steps 6.1-6.3 to achieve dynamic resource scheduling under the evolution of disaster time sequence.

6. A decision-making optimization system for emergency resource scheduling in oilfield disaster scenarios with enhanced temporal characteristics, characterized in that, The system is implemented based on the method of claim 1, and the system includes: The dual-source data processing module is used to configure data parameters in a visual interface, construct synthetic data generation and real data processing dual-source data based on the selected data parameters, and preprocess the dual-source data. The time series feature enhancement module is used to extract the disaster spread speed, impact range, personnel demand, equipment demand, disaster level, and time step characteristics from dual-source data to form an original time series feature matrix. The module then normalizes the original time series feature matrix, converts the normalized feature matrix into a tensor format that the model can input, and outputs a normalized time series feature tensor. The reinforcement learning agent training module is used to input the normalized temporal feature tensor into the TCN causal convolution and self-attention mechanism. The local correlation features of the oilfield disaster time series data are extracted sequentially through the TCN causal convolution layer. The feature range is expanded by an exponential expansion rate, and the extracted local features are weighted and enhanced through the self-attention mechanism to output the enhanced high-dimensional temporal features. The integer programming resource scheduling module is used to construct a reinforcement learning environment for oilfield disaster resource scheduling based on enhanced high-dimensional temporal features, define the state space, action space and two-dimensional reward function, and train the agent model. The dynamic decision optimization fusion module is used to extract the real-time state of the disaster from the enhanced high-dimensional time series features, input it into the intelligent agent model trained in the integer planning resource scheduling module, and output the optimal scheduling action. The integrated visualization monitoring module is used to perform integer planning resource scheduling for the optimal scheduling action, build an integer planning model based on the real-time disaster status, obtain the globally optimal resource scheduling scheme, feed the scheduling results back to the reinforcement learning environment, update the intelligent agent model, output the dynamic resource scheduling scheme, and display the scheduling results through a visualization interface. The method for defining the state space, action space, and two-dimensional reward function in the integer programming resource scheduling module, and training the agent model, is as follows: in This represents the average utilization rate of personnel and equipment resources. ; The average response time for personnel and equipment resources is used, and the DQN and PPO algorithms are employed to train the agent model. The method for extracting the real-time catastrophic state from the enhanced high-dimensional temporal features in the dynamic decision optimization fusion module, inputting it into the agent model trained in step 4, and outputting the optimal scheduling action is as follows: Construct the objective function: in, For personnel response time, For personnel utilization rate, For equipment response time, equipment utilization rate To optimize variables; Set constraints: Total resource constraint: Based on the constructed objective function and constraints, the high-dimensional interior point method is used to solve the integer programming model, and the optimal scheduling personnel, equipment allocation, utilization rate and response time are output.

7. A computer storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-5.

8. A computer device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the method of any one of claims 1-5.

Images