Intelligent unmanned operation and maintenance centralized control method and system for oilfield station based on agent

Abstract

The application discloses an unmanned intelligent operation and maintenance centralized control method and system for oilfield stations based on agents, belongs to the technical field of intelligent operation and maintenance management of oilfield stations, and comprises a data acquisition module, a data summary decision module, a site coordination execution module, an emergency intervention module and an output module.The data acquisition module is used for collecting the crude oil conveying equipment data of interwell stations, transfer stations and joint stations in real time by arranging pressure sensors, temperature sensors, flow sensors, liquid level sensors and water cut sensors in the interwell stations, the transfer stations and the joint stations.The unmanned intelligent operation and maintenance centralized control method and system for oilfield stations based on agents can realize site correlation degree analysis, generate instant optimization adjustment strategies and target optimization adjustment strategies, execute site-specific parameters, feedback and optimize adjustment strategies, record a decision-making process, closely combine a large model with a small model, form a hierarchical control architecture in cooperation with each other, and enhance practicability and functionality.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent operation and maintenance management of oilfield stations, and specifically is a method and system for unmanned intelligent operation and maintenance centralized control of oilfield stations based on intelligent agents. Background Art

[0002] In the current oilfield industry, production efficiency, energy utilization, and safety requirements are constantly increasing. However, traditional manual monitoring and point-to-point control methods are increasingly exposed to their limitations when faced with complex operating conditions of multiple devices. Oilfield stations contain a wide variety of equipment, including but not limited to heaters, pumps, water injection equipment, and free water removers. These devices operate in dynamic and changing conditions during operation, and are highly coupled and complexly correlated with each other.

[0003] Oilfield well stations, transfer stations, and joint stations involve numerous pieces of crude oil transportation equipment, whose operating status is affected by a variety of complex parameters such as pressure, temperature, flow, liquid level, and water content. In the past, manual detection and analysis of the massive amounts of data generated by these devices was difficult to efficiently integrate and accurately analyze. This made it impossible to make timely and reasonable operation and maintenance decisions based on data changes, leading to frequent equipment failures and impacting the overall production efficiency and economic benefits of the oilfield. The inability to immediately identify and adjust abnormal parameters often led to data anomalies at the source site, which simultaneously affected production activities at related sites.

[0004] This case proposes an intelligent agent-based unmanned intelligent operation and maintenance centralized control method and system for oilfield stations to solve the above technical problems. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art. To this end, the present invention proposes an agent-based unmanned intelligent operation and maintenance centralized control method and system for oilfield stations, which solves the above-mentioned technical problems by improving detection and processing methods.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] The intelligent agent-based unmanned intelligent operation and maintenance centralized control system for oilfield stations includes a data acquisition module, a data aggregation and decision-making module, a site coordination and execution module, an emergency intervention module, and an output module.

[0008] The data acquisition module collects data from crude oil transportation equipment at inter-well stations, transfer stations and joint stations in real time by deploying pressure, temperature, flow, liquid level and water content sensors at inter-well stations, transfer stations and joint stations in the oil field;

[0009] The data aggregation and decision-making module receives real-time crude oil transportation equipment data from the inter-well stations, transfer stations, and joint stations under its jurisdiction obtained by the data acquisition module. Based on the data changes of real-time sensors in the inter-well stations and transfer stations, it generates cross-site instant optimization and adjustment strategies for the transfer stations and joint stations through the large model. At the same time, based on the natural language processing technology in the large model, it performs semantic analysis on the data in the oilfield production logs and operation instructions. By combining the analysis model with historical data, it generates target optimization and adjustment strategies for different sites, including combustion control, pump speed adjustment, and water dilution ratio optimization, and transmits them to the site coordination and execution module.

[0010] The site coordination execution module, based on the immediate optimization adjustment strategy and target optimization adjustment strategy generated for each site by the data aggregation decision module, combines the data optimization small models deployed at each site for deep optimization and fine control of specific equipment and functions, and executes corresponding adjustment strategies for each site respectively. At the same time, based on the data changes of the small models of different sites during the execution process, the large model optimization adjustment strategy is further fine-tuned according to real-time monitoring data.

[0011] Furthermore, when the data aggregation decision module generates the immediate optimization adjustment strategy and the target optimization adjustment strategy through the large model, the following steps are included:

[0012] The data acquisition module receives real-time pressure, temperature, flow, liquid level, and water content data from subordinate inter-well stations and transfer stations. Based on abnormal changes in real-time data, it identifies abnormal inter-well stations and transfer stations. Based on the crude oil transportation relationship network of different abnormal inter-well stations and transfer stations, it generates cross-site instant optimization and adjustment strategies for related transfer stations and joint stations through the LMM large model.

[0013] Based on the natural language processing technology in the large model, semantic analysis is performed on the data in the oilfield production logs and updated operation instructions in the well stations, transfer stations, and joint stations. According to the analysis model and combined with historical data, target optimization adjustment strategies for different sites are generated, including combustion control, pump speed adjustment, and water dilution ratio optimization, and transmitted to the site coordination execution module.

[0014] Furthermore, the specific steps for constructing the LMM large model for the subordinate well stations, transfer stations, and joint stations are as follows:

[0015] The statistical system manages information on inter-well stations, transfer stations, and joint stations, including station locations and upstream and downstream relationships. A network of inter-well stations, transfer stations, and joint stations is established. Historical sensor data from different stations is collected, including pressure data, temperature data, flow data, liquid level data, and crude oil water content data. The mean μ and standard value σ of the historical sensor data corresponding to equipment at different stations are calculated, and the normal range (μ-2σ, μ+2σ) of sensor data at different stations is set according to the 3sigma principle.

[0016] Based on the normal intervals of sensor data at different sites, the abnormal parameter conditions of sensor data at each site are recorded separately, including those below the normal interval and those above the normal interval, which are recorded as low anomalies and high anomalies respectively. The adjustment strategies of related sites in the relationship network under the corresponding abnormal parameter conditions are collected, and the data of upstream and downstream sites of the abnormal site in the same time window are extracted to form a set of related adjustment strategy events. The abnormal parameter conditions and the adjustment strategies of related sites are combined into training samples. Each sample contains input and output. An open source large language model is selected as the framework, and specific modifications are made according to the characteristics of oilfield site data. An input layer is added to process structured abnormal parameter information, and an output layer is set to generate the adjustment strategy text under the corresponding input layer. The constructed training samples are input into the model in batches, and the model output is calculated by forward propagation. The loss value is calculated according to the loss function, and the model parameters are updated through the Adam optimization algorithm. Multiple rounds are repeated until the model converges or the preset number of training rounds is reached.

[0017] Furthermore, based on abnormal changes in real-time data, abnormal well stations and transfer stations are identified. According to the crude oil transportation relationship network of different abnormal well stations and transfer stations, the LMM large model is used to generate cross-site instant optimization adjustment strategies for related transfer stations and joint stations, including the following steps:

[0018] The data acquisition module receives the real-time pressure, temperature, flow, liquid level and water content data of the subordinate well stations and transfer stations, and calculates the mean according to the time window. Based on the normal interval of the sensor data of different stations, the mean data not included in the interval is judged. When the sensor mean data is less than the normal interval of the current sensor data of the corresponding station, the current sensor data of the current station is judged to be low abnormal. When the sensor mean data is higher than the normal interval of the current sensor data of the corresponding station, the current sensor data of the current station is judged to be high abnormal.

[0019] The abnormal parameters of abnormal well stations and transfer stations and their location information in the relationship network are one-hot encoded. The LMM large model is used to generate cross-site instant optimization adjustment strategies for related transfer stations and joint stations. The probability of each adjustment strategy is obtained through the softmax function, and the instant optimization adjustment strategy with the highest probability is transmitted to the site coordination execution module.

[0020] Furthermore, the natural language processing technology in the large model is used to generate target optimization adjustment strategies for different sites, including the following steps:

[0021] Production logs and operation instruction data, including operation scenarios and equipment information, are collected. The BiLSTM-CRF model is trained using the labeled production logs and operation instruction data. The operation instructions are analyzed using a dependency syntax analysis tool. A dependency syntax tree is constructed to clarify the dependency relationships between words. Based on the dependency syntax tree and the identified entities, the causes, actions, and target parameters in the causal relationship are extracted to form structured triples (triple = (cause, action, target parameter). Based on the generated triples, a target optimization and adjustment strategy suitable for the subordinate sites is matched and transmitted to the site coordination execution module.

[0022] Furthermore, the instant optimization adjustment strategy and target optimization adjustment strategy generated by the data aggregation decision module for each site are respectively executed on each site through the small model, including the following steps:

[0023] For different sites under its jurisdiction, based on the historical data of the equipment in each site, a data adjustment prediction model for the corresponding equipment is established through a neural network. According to the immediate optimization adjustment strategy and the target optimization adjustment strategy transmitted by the data summary decision module, the optimization adjustment sites and optimization adjustment equipment involved are determined. Based on the parameter fluctuation value of the influencing source, the data adjustment value is calculated through the data adjustment prediction model of the corresponding equipment at the site involved. The specific steps are as follows:

[0024] For different equipment in different well stations, transfer stations, and joint stations, the long short-term memory network is used with the mean square error as the loss function combined with the Adam optimizer to build the equipment parameter adjustment prediction model for the corresponding station. Based on the real-time optimization adjustment strategy and the target optimization adjustment strategy of the data summary decision module, the prediction model is constructed. Perform policy device matching, where v 策略 、v 设备 They are the strategy feature vector and the device history data feature vector respectively;

[0025] Determine the source of fluctuation of key parameters through sensitivity analysis and calculate the parameter x j The gradient of the target Y, where:

[0026]

[0027] Among them, h k Represents the output value of the kth hidden layer in the neural network, Y is the optimization target variable, x j is the device parameter, L represents the total number of hidden layers in the neural network;

[0028] Set the strategic goal Y 目标 And the current parameter x is input into the model to obtain the parameter adjustment Δx, combined with the equipment safety threshold, where the input model is:

[0029]

[0030] Among them, Δx is the parameter adjustment amount, that is, the device parameter change value that needs to be calculated, is the neural network prediction model, that is, the target value predicted after the input parameter x, λ represents the regularization coefficient, and then constructs a constrained optimization problem:

[0031] minimize Satisfies x+Δx≤x max ,x+Δx≥x min ;

[0032] Among them, x min 、x max They represent the physical safety thresholds of equipment parameters respectively. The Lagrange multiplier method is used to finally obtain the optimal adjustment value, and the parameters are adjusted at the corresponding site based on the optimal adjustment value.

[0033] Furthermore, the optimization and adjustment strategy of the large model is further fine-tuned based on the data changes during the execution of the small models of different sites according to the real-time monitoring data, including the following steps:

[0034] During the parameter adjustment process of the small model, the real-time data of the small model at each site will be uploaded to the central data pool. Calculate the deviation δ between the actual value and the target value, Represent the actual value and the target value respectively, and decompose the deviation into the contribution of each associated site through the chain rule:

[0035]

[0036] Among them, y i represents the key parameter of the i-th site, Δy i The actual adjustment amount is updated based on real-time feedback data. The policy value function Q(s,a) of the large model is updated, the influence matrix between sites is defined, and the constrained optimization problem is reconstructed. At the same time, the adjusted policy parameters are verified to be within the safety threshold. The policy parameter f min ≤f 新 ≤f max ,If the strategy parameters are out of bounds, the projection method is used for correction;

[0037] At the same time, the strategy parameters and corresponding performance indicators after each adjustment are recorded to build a historical version library and re-solve the global optimization problem every fixed time window.

[0038] Furthermore, in the emergency intervention module, in the data aggregation decision module and the site coordination control module, when an error occurs in the model, the system will automatically generate an event report and notify the on-duty operation and maintenance personnel via SMS or APP notification. The generated event report content includes the error code and number, the name of the sensor involved, and the site information involved in the error;

[0039] The output module uses the initial state of the equipment before adjustment and related environmental data as the starting point of reasoning after each data aggregation decision module and the site coordination execution module perform site equipment parameter adjustment optimization. It records the rules and algorithms based on which the data aggregation decision module formulates the optimization strategy, and each step in the process of executing the optimization strategy by the site coordination execution module, including the order of parameter adjustment, the adjustment range, and the feedback data of the equipment after each adjustment, and outputs a complete and detailed text description of the optimization strategy reasoning process data.

[0040] The agent-based unmanned intelligent operation and maintenance centralized control method for oilfield stations includes the following steps:

[0041] S1. By deploying pressure, temperature, flow, liquid level, and water content sensors at oilfield inter-well stations, transfer stations, and joint stations, real-time data on crude oil transportation equipment at these stations is collected. The data is received from subordinate inter-well stations, transfer stations, and joint stations in real time. Based on the data changes from real-time sensors at these stations, the LMM large model is used to generate cross-site instant optimization and adjustment strategies for transfer stations and joint stations, as well as target optimization and adjustment strategies.

[0042] S2: Based on the real-time optimization and target optimization strategies for each site, and in combination with the data optimization models deployed at each site for deep optimization and fine-grained control of specific equipment and functions, the corresponding adjustment strategies are implemented for each site. Furthermore, based on data changes during the execution of the small models at different sites, the large-scale model optimization and adjustment strategies are further fine-tuned based on real-time monitoring data.

[0043] S3. When the model reports an error, it automatically generates an event report and notifies the on-duty operation and maintenance personnel via SMS or APP notification. The generated event report includes the error code and number, the name of the sensor involved, and the site information involved in the error.

[0044] S4. After each site equipment parameter adjustment and optimization, use the initial state of the equipment before adjustment and related environmental data as the starting point for reasoning, record the rules and algorithms based on which the optimization strategy is formulated, and record each step in the site's execution of the optimization strategy, including the order of parameter adjustment, the magnitude of adjustment, and the feedback data from the equipment after each adjustment. Output a complete text description of the optimization strategy reasoning process data.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. In this invention, by collecting statistics on oilfield station types and based on the correlation between different stations, a hierarchical decision-making system of large models + small models is used to integrate, analyze and make decisions on the data. This enables the overall automated management of oilfield stations from daily operation and maintenance to emergency management. Based on the real-time sensor data changes at well stations and transfer stations, cross-site instant optimization and adjustment strategies are quickly generated. At the same time, target optimization and adjustment strategies are generated in combination with historical data, thereby improving the overall operational efficiency of oilfield stations in unmanned conditions.

[0047] 2. In this invention, through the cooperation of the site coordination execution module and the data aggregation decision module, the small model is optimized in combination with the data of each site, and the site macro adjustment strategy is refined. The large model strategy is fine-tuned according to the data changes during the execution of the small model. The continuous optimization of key parameters such as combustion control, pump speed adjustment, and water dilution ratio in the oil field production process is achieved, so as to improve the actual crude oil production quality under different circumstances.

[0048] 3. In the present invention, the output module records the optimization strategy reasoning process of each large model in conjunction with the small model, including the rule algorithm based on the decision, the execution steps and the equipment feedback data, etc., providing complete and accurate data support for subsequent analysis and summary as well as problem tracing, making it easier for managers to understand the actual situation in the operation and maintenance production process and facilitate continuous improvement of operation and maintenance management work;

[0049] The entire intelligent agent-based unmanned intelligent operation and maintenance centralized control method and system for oilfield stations can realize site correlation analysis, real-time optimization adjustment strategy and target optimization adjustment strategy generation, site-specific parameter execution, feedback and optimization adjustment strategy fine-tuning, and decision-making process recording. It closely combines large models with small models to form a hierarchical control architecture with top-down coordination, enhancing practicality and functionality. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a block diagram of the agent-based unmanned intelligent operation and maintenance centralized control system for oilfield stations of the present invention;

[0051] Figure 2 This is a flow chart of the agent-based unmanned intelligent operation and maintenance centralized control method for oilfield stations of the present invention. DETAILED DESCRIPTION

[0052] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0053] Example 1:

[0054] like Figure 1 As shown in the figure, the intelligent agent-based unmanned intelligent operation and maintenance centralized control system for oilfield stations includes a data acquisition module, a data aggregation and decision-making module, a site coordination and execution module, an emergency intervention module, and an output module.

[0055] The data acquisition module collects real-time data from crude oil transportation equipment at oilfield well stations, transfer stations, and joint stations by deploying pressure, temperature, flow, liquid level, and water content sensors at these stations.

[0056] In the emergency intervention module, when a model error occurs in the data aggregation and decision-making module and the site coordination and control module, the system automatically generates an event report and notifies the on-duty operation and maintenance personnel via SMS and APP notification. The generated event report includes the error code and number, the name of the sensor involved, and the site information involved in the error.

[0057] The output module, after each time the data aggregation and decision module and the site coordination and execution module perform site equipment parameter adjustment and optimization, uses the initial state of the equipment before adjustment and related environmental data as the starting point for reasoning. It records the rules and algorithms based on which the data aggregation and decision module formulates the optimization strategy, and records each step in the site coordination and execution module's execution of the optimization strategy, including the order and magnitude of parameter adjustments, as well as the feedback data from the equipment after each adjustment. It then outputs a detailed textual description of the complete optimization strategy reasoning process data.

[0058] The data aggregation and decision-making module receives real-time crude oil transportation equipment data from the inter-well stations, transfer stations, and joint stations under its jurisdiction obtained by the data acquisition module. Based on the data changes of real-time sensors in the inter-well stations and transfer stations, it generates cross-site instant optimization and adjustment strategies for transfer stations and joint stations through the large model. At the same time, based on the natural language processing technology in the large model, it performs semantic analysis on the data in the oilfield production logs and operation instructions. By combining the analysis model with historical data, it generates target optimization and adjustment strategies for different sites, including combustion control, pump speed adjustment, and water dilution ratio optimization, and transmits them to the site coordination and execution module.

[0059] The specific steps for constructing the LMM large model for the subordinate well stations, transfer stations, and joint stations are as follows:

[0060] The statistical system is under the jurisdiction of interwell stations, transfer stations and joint stations, including station location, upstream and downstream relationship of the station, establishing the relationship network of interwell stations, transfer stations and joint stations, collecting historical sensor data of different stations, including pressure data, temperature data, flow data, liquid level data and crude oil water content data, calculating the mean value μ and standard value σ of the historical sensor data of the corresponding equipment at different stations, setting the normal interval (μ-2σ, μ+2σ) of the sensor data of different stations through the 3 sigma principle;

[0061] It should be noted that the interwell station, transfer station and joint station are modeled as a directed graph G=(V, E), where V and E represent the station and upstream and downstream flow relationship, respectively, and an adjacency topology matrix A∈{0,1} N×N is constructed, where if there is a direct upstream and downstream relationship between station i and station j, then A ij =1, otherwise A ij =0, where N is the current number of stations, and for interwell stations, transfer stations and joint stations, pressure data, temperature data, flow data, liquid level data and crude oil water content data of the corresponding equipment at the station are collected, including heating furnace combustion chamber temperature, outlet crude oil temperature, ambient temperature, gas pipeline pressure, combustion chamber pressure, gas supply amount, crude oil flow, pump inlet pressure, pump outlet pressure, water mixing pipeline pressure, mixed liquid pipeline pressure, pump liquid flow, crude oil water content, water mixing flow, mixed liquid total flow, storage tank liquid level, buffer tank liquid level, etc. to perform real-time optimization and adjustment through a large model.

[0062] Based on the normal interval of different station sensor data, the abnormal parameter conditions of each station sensor data are recorded, including below the normal interval and above the normal interval, which are recorded as low abnormality and high abnormality, respectively, and the adjustment strategy of the associated station in the relationship network under the corresponding abnormal parameter condition is collected, the data of the upstream and downstream stations of the abnormal station in the same time window is extracted to form an associated adjustment strategy event set, the abnormal parameter condition and the adjustment strategy of the associated station are combined into a training sample, each sample contains input and output, an open source large language model is selected as a framework, and specific modifications are made according to the characteristics of the oilfield station data, an input layer is added to process structured abnormal parameter information, and an output layer is set to generate adjustment strategy text under the corresponding input layer. The trained samples are input into the model in batches, the forward propagation calculation model output is calculated, the loss value is calculated according to the loss function, the model parameters are updated through the Adam optimization algorithm, and the process is repeated for multiple rounds until the model converges or the preset training rounds are reached.

[0063] It should be noted that the input of each sample is an abnormal parameter situation, such as abnormally high pressure and abnormally low temperature at well station A. At the same time, the abnormal parameters are converted into natural language descriptions, such as "The pressure at well station A is higher than the normal range" as a text prefix. Using the Seq2Seq structure, the output of each sample is the adjustment strategy of the associated site, such as increasing the valve opening by 10% at transfer station B and reducing the heating power by 10% at joint station C. The input layer encodes the abnormal parameter situation into a vector form through one-hot encoding and inputs it into the model. For the classification task of the adjustment strategy, different types of adjustment strategies are regarded as different categories, and the cross-entropy loss function is used as the loss function. Based on the special adjustment behavior of oilfield sites, the loss value is calculated by combining the mean square error loss with the cross-entropy loss function for the numerical parameters in the adjustment strategy, such as the valve opening adjustment ratio. The weight ratio is set to 0.5 and 0.5, respectively. The ratio of the training set, validation set, and test set is 70%, 15%, and 15%. Open source large language models include GPT-Neo and Deepseek.

[0064] When the data aggregation decision module generates the immediate optimization adjustment strategy and the target optimization adjustment strategy through the large model, the following steps are included:

[0065] The data acquisition module receives the real-time pressure, temperature, flow, liquid level, and water content data of the subordinate inter-well stations and transfer stations. Based on the abnormal changes in the real-time data, abnormal inter-well stations and transfer stations are identified. Based on the crude oil transportation relationship network of different abnormal inter-well stations and transfer stations, the LMM large model is used to generate cross-site instant optimization adjustment strategies for related transfer stations and joint stations, including the following steps:

[0066] The data acquisition module receives the real-time pressure, temperature, flow, liquid level and water content data of the subordinate well stations and transfer stations, and calculates the mean according to the time window. Based on the normal interval of the sensor data of different stations, the mean data not included in the interval is judged. When the sensor mean data is less than the normal interval of the current sensor data of the corresponding station, the current sensor data of the current station is judged to be low abnormal. When the sensor mean data is higher than the normal interval of the current sensor data of the corresponding station, the current sensor data of the current station is judged to be high abnormal.

[0067] The abnormal parameters of abnormal well stations and transfer stations and their location information in the relationship network are one-hot encoded. The LMM large model is used to generate cross-site instant optimization adjustment strategies for related transfer stations and joint stations. The probability of each adjustment strategy is obtained through the softmax function, and the instant optimization adjustment strategy with the highest probability is transmitted to the site coordination execution module.

[0068] It should be noted that the macro-adjustment strategy of the associated sites is obtained based on the LMM large model. For example, if the input information is that the pressure of well station A is higher than the normal range, based on the relationship network between well station A and transfer station B, the output result is to increase the valve opening of transfer station B. The output result is transmitted to the site coordination execution module for further parameter adjustment within the site.

[0069] Based on the natural language processing technology in the large model, semantic analysis is performed on the data in the oilfield production logs and update operation instructions in the well station, transfer station, and joint station. Based on the analysis model and historical data, target optimization adjustment strategies for different stations are generated, including combustion control, pump speed adjustment, and water dilution ratio optimization. These strategies are then transmitted to the station coordination execution module, including the following steps:

[0070] Production logs and operation instruction data, including operation scenarios and equipment information, are collected. The BiLSTM-CRF model is trained using the labeled production logs and operation instruction data. The operation instructions are analyzed using a dependency syntax analysis tool. A dependency syntax tree is constructed to clarify the dependency relationships between words. Based on the dependency syntax tree and the identified entities, the causes, actions, and target parameters in the causal relationship are extracted to form structured triples (triple = (cause, action, target parameter). Based on the generated triples, a target optimization and adjustment strategy suitable for the subordinate sites is matched and transmitted to the site coordination execution module.

[0071] It should be noted that the preprocessed production log is input into the trained NER model to identify key parameters, operating actions and equipment identifications, among which key parameters include temperature, pressure, flow, etc., operating actions include opening, closing, adjusting, etc., and equipment identifications include pumps, valves, burners, etc. According to the generated triples and pre-set rules, the target optimization adjustment strategy suitable for the subordinate sites is matched. For example, triple = (temperature is too high, reduce, combustion power), then the target optimization adjustment strategy generated according to the relevant rules is the combustion control strategy, and the combustion control strategy signal is transmitted into the site coordination execution module, and further parameter refinement is performed through the small model.

[0072] Example 2:

[0073] The site coordination execution module, based on the real-time optimization adjustment strategy and target optimization adjustment strategy generated by the data aggregation decision module for each site, combines the data optimization small models deployed at each site for deep optimization and fine-grained control of specific equipment and functions, and executes the corresponding adjustment strategy for each site. At the same time, based on the data changes during the execution of the small models at different sites, the large model optimization adjustment strategy is further fine-tuned according to real-time monitoring data. The steps include:

[0074] For different sites under its jurisdiction, based on the historical data of the equipment in each site, a data adjustment prediction model for the corresponding equipment is established through a neural network. According to the immediate optimization adjustment strategy and the target optimization adjustment strategy transmitted by the data summary decision module, the optimization adjustment sites and optimization adjustment equipment involved are determined. Based on the parameter fluctuation value of the influencing source, the data adjustment value is calculated through the data adjustment prediction model of the corresponding equipment of the site involved. The specific steps are as follows:

[0075] For different equipment in different well stations, transfer stations, and joint stations, the long short-term memory network is used with the mean square error as the loss function combined with the Adam optimizer to build the equipment parameter adjustment prediction model for the corresponding station. Based on the real-time optimization adjustment strategy and the target optimization adjustment strategy of the data summary decision module, the prediction model is constructed. Perform policy device matching, where v 策略 、v 设备 They are the strategy feature vector and the device history data feature vector respectively;

[0076] It should be noted that v 策略 Represents the strategy feature vector, which represents the mathematical characteristics of the optimization strategy. The vector dimensions include energy consumption, efficiency, safety, adjustment range, priority weight, etc. For example, when the optimization strategy is "reducing the energy consumption of the well station", the vector is represented as [energy consumption weight, adjustment range, equipment type code], v 设备 is the equipment feature vector, which represents the mathematical characteristics of the equipment's historical data and functions, including parameters such as equipment type, historical energy consumption, failure frequency, and operating efficiency. For example, the feature vector of a pump is expressed as [mean energy consumption, maximum flow rate, vibration threshold]. The closer cos(θ) is to 1, the higher the match between the current optimization strategy and the current equipment.

[0077] Determine the source of fluctuation of key parameters through sensitivity analysis and calculate the parameter x j The gradient of the target Y, where:

[0078]

[0079] Among them, h k represents the output value of the kth hidden layer in the neural network, Y is the optimization target variable, x j is the device parameter, L represents the total number of hidden layers in the neural network;

[0080] It should be noted that the optimization target variables include energy consumption, flow rate, safety score, etc., and equipment parameters include pump speed, combustion temperature, and water mixing ratio. By calculating the gradient of the target to the parameter, the sensitivity of the parameter to the target can be quantified. The larger the absolute value of the gradient, the more effective the adjustment is in achieving the target.

[0081] Set the strategic goal Y目标 And the current parameter x is input into the model to obtain the parameter adjustment Δx, combined with the equipment safety threshold, where the input model is:

[0082]

[0083] Among them, Δx is the parameter adjustment amount, that is, the device parameter change value that needs to be calculated, is the neural network prediction model, that is, the target value predicted after the input parameter x, λ represents the regularization coefficient, and then constructs a constrained optimization problem:

[0084] minimize Satisfies x+Δx≤x max ,x+Δx≥x min ;

[0085] Among them, x min 、x max They represent the physical safety thresholds of equipment parameters respectively. The Lagrange multiplier method is used to finally obtain the optimal adjustment value, and the parameters are adjusted at the corresponding site based on the optimal adjustment value.

[0086] Based on the data changes during the execution of the small models at different sites, the optimization and adjustment strategy of the large model is further fine-tuned according to real-time monitoring data, including the following steps:

[0087] During the parameter adjustment process of the small model, the real-time data of the small model at each site will be uploaded to the central data pool. Calculate the deviation δ between the actual value and the target value, Represent the actual value and the target value respectively, and decompose the deviation into the contribution of each associated site through the chain rule:

[0088]

[0089] Among them, y i represents the key parameter of the i-th site, Δy i The actual adjustment amount is updated based on real-time feedback data. The policy value function Q(s,a) of the large model is updated, the influence matrix between sites is defined, and the constrained optimization problem is reconstructed. At the same time, the adjusted policy parameters are verified to be within the safety threshold. The policy parameter f min ≤f 新 ≤f max ,If the strategy parameters are out of bounds, the projection method is used for correction;

[0090] It should be noted that:

[0091]

[0092] Update the strategy value function of the large model, s, a represent the state and action respectively, s′, a′ represent the transfer state and action after execution, α, γ, r represent the learning rate, discount factor, and immediate reward respectively. The update step size is controlled by the learning rate, and the current and future rewards are balanced by the discount factor. The immediate rewards include the energy consumption reduction, water ratio optimization parameters, combustion control parameters, etc. The influence matrix P between sites is defined, where the element p nm represents the coupling coefficient of site m to site n. The global objective function is updated to include the multi-site coupling effect:

[0093]

[0094] Where M represents the total number of sites. Represents the coupling penalty coefficient to suppress excessive cross-site interference. When the parameter exceeds the limit, f is obtained by projection method. 修正 :

[0095]

[0096] At the same time, the strategy parameters and corresponding performance indicators after each adjustment are recorded to build a historical version library and re-solve the global optimization problem every fixed time window.

[0097] It should be noted that the global optimization problem is solved again every fixed time window. The specific steps are as follows:

[0098]

[0099] Among them, N represents the number of data points in the window, Y t Represents the actual value at time point t.

[0100] Example 3:

[0101] like Figure 2 As shown in FIG, the agent-based unmanned intelligent operation and maintenance centralized control method for oilfield stations includes the following steps:

[0102] S1. By deploying pressure, temperature, flow, liquid level, and water content sensors at oilfield inter-well stations, transfer stations, and joint stations, real-time data on crude oil transportation equipment at these stations is collected. The data is received from subordinate inter-well stations, transfer stations, and joint stations in real time. Based on the data changes from real-time sensors at these stations, the LMM large model is used to generate cross-site instant optimization and adjustment strategies for transfer stations and joint stations, as well as target optimization and adjustment strategies.

[0103] S2: Based on the real-time optimization and target optimization strategies for each site, and in combination with the data optimization models deployed at each site for deep optimization and fine-grained control of specific equipment and functions, the corresponding adjustment strategies are implemented for each site. Furthermore, based on data changes during the execution of the small models at different sites, the large-scale model optimization and adjustment strategies are further fine-tuned based on real-time monitoring data.

[0104] S3. When the model reports an error, it automatically generates an event report and notifies the on-duty operation and maintenance personnel via SMS or APP notification. The generated event report includes the error code and number, the name of the sensor involved, and the site information involved in the error.

[0105] S4. After each site equipment parameter adjustment and optimization, use the initial state of the equipment before adjustment and related environmental data as the starting point for reasoning, record the rules and algorithms based on which the optimization strategy is formulated, and record each step in the site's execution of the optimization strategy, including the order of parameter adjustment, the magnitude of adjustment, and the feedback data from the equipment after each adjustment. Output a complete text description of the optimization strategy reasoning process data.

[0106] The present invention is based on an intelligent agent-based unmanned intelligent operation and maintenance centralized control method and system for oilfield stations. When in use, by statistically analyzing the types of oilfield stations and based on the correlation between different sites, a hierarchical decision-making system of large models + small models is used to fuse, analyze and make decisions on the data, thereby realizing overall automated management of oilfield stations from daily operation and maintenance to emergency management. Based on the real-time sensor data changes of well stations and transfer stations, cross-site instant optimization and adjustment strategies are quickly generated. At the same time, target optimization and adjustment strategies are generated in combination with historical data, thereby improving the overall operation efficiency of oilfield stations in unmanned conditions.

[0107] In the embodiments provided by the present invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the module division is merely a logical functional division, and other division methods may be used in actual implementation. The modules described as separate components may or may not be physically separate, and the components displayed as modules may or may not be physical units, that is, they may be located in one place or distributed across multiple network elements. Some or all of the modules may be selected according to actual needs to achieve the purpose of the method of this embodiment.

[0108] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. An agent-based unmanned intelligent operation and maintenance centralized control system for oilfield stations, characterized by: It includes data collection module, data summary and decision module, site coordination and execution module, emergency intervention module and output module; The data acquisition module collects data from crude oil transportation equipment at inter-well stations, transfer stations and joint stations in real time by deploying pressure, temperature, flow, liquid level and water content sensors at inter-well stations, transfer stations and joint stations in the oil field; The data aggregation and decision-making module receives real-time crude oil transportation equipment data from the inter-well stations, transfer stations, and joint stations under its jurisdiction obtained by the data acquisition module. Based on the data changes of real-time sensors in the inter-well stations and transfer stations, it generates cross-site instant optimization and adjustment strategies for the transfer stations and joint stations through the large model. At the same time, based on the natural language processing technology in the large model, it performs semantic analysis on the data in the oilfield production logs and operation instructions. By combining the analysis model with historical data, it generates target optimization and adjustment strategies for different sites, including combustion control, pump speed adjustment, and water dilution ratio optimization, and transmits them to the site coordination and execution module. The site coordination execution module, based on the real-time optimization adjustment strategy and target optimization adjustment strategy generated by the data aggregation decision module for each site, combines the data optimization small models deployed at each site for deep optimization and fine control of specific equipment and functions, and executes the corresponding adjustment strategy for each site. At the same time, based on the data changes during the execution of the small models at different sites, the large model optimization adjustment strategy is further fine-tuned according to real-time monitoring data; Based on the data changes during the execution of the small models at different sites, the optimization and adjustment strategy of the large model is further fine-tuned according to real-time monitoring data, including the following steps: During the parameter adjustment process of the small model, the real-time data of the small model at each site will be uploaded to the central data pool. Calculate the deviation δ between the actual value and the target value, Represent the actual value and the target value respectively, and decompose the deviation into the contribution of each associated site through the chain rule: Among them, y i represents the key parameter of the i-th site, Δy i The actual adjustment amount is updated based on real-time feedback data. The policy value function Q(s,a) of the large model is updated, the influence matrix between sites is defined, and the constrained optimization problem is reconstructed. At the same time, the adjusted policy parameters are verified to be within the safety threshold. The policy parameter f min ≤f 新 ≤f max ,If the strategy parameters are out of bounds, the projection method is used for correction; pass: Update the strategy value function of the large model, s, a represent the state and action respectively, s′, a′ represent the transfer state and action after execution, α, γ, r represent the learning rate, discount factor, and immediate reward respectively. The update step size is controlled by the learning rate, and the current and future rewards are balanced by the discount factor. The immediate reward includes the energy consumption reduction, water ratio optimization parameters, and combustion control parameters. Define the influence matrix P between sites, where the element p nm Represents the coupling coefficient of site m to site n. The global objective function is updated to include the multi-site coupling effect: Where M represents the total number of sites. Represents the coupling penalty coefficient to suppress excessive cross-site interference. When the parameter exceeds the limit, f is obtained by projection method. 修正 : At the same time, the strategy parameters and corresponding performance indicators after each adjustment are recorded to build a historical version library and re-solve the global optimization problem every fixed time window.

2. The agent-based unmanned intelligent operation and maintenance centralized control system for oilfield stations according to claim 1 is characterized by: When the data aggregation decision module generates the immediate optimization adjustment strategy and the target optimization adjustment strategy through the large model, the following steps are included: The data acquisition module receives real-time pressure, temperature, flow, liquid level, and water content data from subordinate inter-well stations and transfer stations. Based on abnormal changes in real-time data, it identifies abnormal inter-well stations and transfer stations. Based on the crude oil transportation relationship network of different abnormal inter-well stations and transfer stations, it generates cross-site instant optimization and adjustment strategies for related transfer stations and joint stations through the LMM large model. Based on the natural language processing technology in the large model, semantic analysis is performed on the data in the oilfield production logs and updated operation instructions in the well stations, transfer stations, and joint stations. According to the analysis model and combined with historical data, target optimization adjustment strategies for different sites are generated, including combustion control, pump speed adjustment, and water dilution ratio optimization, and transmitted to the site coordination execution module.

3. The agent-based unmanned intelligent operation and maintenance centralized control system for oilfield stations according to claim 2 is characterized by: The specific steps for constructing the LMM large model for the subordinate well stations, transfer stations, and joint stations are as follows: The statistical system manages information on inter-well stations, transfer stations, and joint stations, including station locations and upstream and downstream relationships. A network of inter-well stations, transfer stations, and joint stations is established. Historical sensor data from different stations is collected, including pressure data, temperature data, flow data, liquid level data, and crude oil water content data. The mean μ and standard value σ of the historical sensor data corresponding to equipment at different stations are calculated, and the normal range (μ-2σ, μ+2σ) of sensor data at different stations is set according to the 3sigma principle. Based on the normal intervals of sensor data at different sites, the abnormal parameter conditions of sensor data at each site are recorded separately, including those below the normal interval and those above the normal interval, which are recorded as low anomalies and high anomalies respectively. The adjustment strategies of related sites in the relationship network under the corresponding abnormal parameter conditions are collected, and the data of upstream and downstream sites of the abnormal site in the same time window are extracted to form a set of related adjustment strategy events. The abnormal parameter conditions and the adjustment strategies of related sites are combined into training samples. Each sample contains input and output. An open source large language model is selected as the framework, and specific modifications are made according to the characteristics of oilfield site data. An input layer is added to process structured abnormal parameter information, and an output layer is set to generate the adjustment strategy text under the corresponding input layer. The constructed training samples are input into the model in batches, and the model output is calculated by forward propagation. The loss value is calculated according to the loss function, and the model parameters are updated through the Adam optimization algorithm. Multiple rounds are repeated until the model converges or the preset number of training rounds is reached.

4. The agent-based unmanned intelligent operation and maintenance centralized control system for oilfield stations according to claim 3 is characterized by: Based on abnormal changes in real-time data, abnormal well stations and transfer stations are identified. Based on the crude oil transportation relationship network of different abnormal well stations and transfer stations, the LMM large model is used to generate cross-site instant optimization and adjustment strategies for related transfer stations and joint stations. The following steps are included: The data acquisition module receives the real-time pressure, temperature, flow, liquid level and water content data of the subordinate well stations and transfer stations, and calculates the mean according to the time window. Based on the normal interval of the sensor data of different stations, the mean data not included in the interval is judged. When the sensor mean data is less than the normal interval of the current sensor data of the corresponding station, the current sensor data of the current station is judged to be low abnormal. When the sensor mean data is higher than the normal interval of the current sensor data of the corresponding station, the current sensor data of the current station is judged to be high abnormal. The abnormal parameters of abnormal well stations and transfer stations and their location information in the relationship network are one-hot encoded. The LMM large model is used to generate cross-site instant optimization adjustment strategies for related transfer stations and joint stations. The probability of each adjustment strategy is obtained through the softmax function, and the instant optimization adjustment strategy with the highest probability is transmitted to the site coordination execution module.

5. The agent-based unmanned intelligent operation and maintenance centralized control system for oilfield stations according to claim 3 is characterized by: The method of generating target optimization adjustment strategies for different sites based on natural language processing technology in the large model includes the following steps: Production logs and operation instruction data, including operation scenarios and equipment information, are collected. The BiLSTM-CRF model is trained using the labeled production logs and operation instruction data. The operation instructions are analyzed using a dependency syntax analysis tool. A dependency syntax tree is constructed to clarify the dependency relationships between words. Based on the dependency syntax tree and the identified entities, the causes, actions, and target parameters in the causal relationship are extracted to form structured triples (triple = (cause, action, target parameter). Based on the generated triples, a target optimization and adjustment strategy suitable for the subordinate sites is matched and transmitted to the site coordination execution module.

6. The agent-based unmanned intelligent operation and maintenance centralized control system for oilfield stations according to claim 1 is characterized by: The instant optimization adjustment strategy and target optimization adjustment strategy generated by the data aggregation decision module for each site are respectively executed on each site through the small model, including the following steps: For different sites under its jurisdiction, based on the historical data of the equipment in each site, a data adjustment prediction model for the corresponding equipment is established through a neural network. According to the immediate optimization adjustment strategy and the target optimization adjustment strategy transmitted by the data summary decision module, the optimization adjustment sites and optimization adjustment equipment involved are determined. Based on the parameter fluctuation value of the influencing source, the data adjustment value is calculated through the data adjustment prediction model of the corresponding equipment at the site involved. The specific steps are as follows: For different equipment in different well stations, transfer stations, and joint stations, the long short-term memory network is used with the mean square error as the loss function combined with the Adam optimizer to build the equipment parameter adjustment prediction model for the corresponding station. Based on the real-time optimization adjustment strategy and the target optimization adjustment strategy of the data summary decision module, the prediction model is constructed. Perform policy device matching, where v 策略 、v 设备 They are the strategy feature vector and the device history data feature vector respectively; Determine the source of fluctuation of key parameters through sensitivity analysis and calculate the parameter x j The gradient of the target Y, where: Among them, h k Represents the output value of the kth hidden layer in the neural network, Y is the optimization target variable, x j is the device parameter, L represents the total number of hidden layers in the neural network; Set the strategic goal Y 目标 And the current parameter x is input into the model to obtain the parameter adjustment Δx, combined with the equipment safety threshold, where the input model is: Among them, Δx is the parameter adjustment amount, that is, the device parameter change value that needs to be calculated, is the neural network prediction model, that is, the target value predicted after the input parameter x, λ represents the regularization coefficient, and then constructs a constrained optimization problem: minimize Satisfies x+Δx≤x max ,x+Δx≥x min ; Among them, x min 、x max They represent the physical safety thresholds of equipment parameters respectively. The Lagrange multiplier method is used to finally obtain the optimal adjustment value, and the parameters are adjusted at the corresponding site based on the optimal adjustment value.

7. The agent-based unmanned intelligent operation and maintenance centralized control system for oilfield stations according to claim 1 is characterized by: The emergency intervention module, in the data aggregation decision module and the site coordination control module, when the model reports an error, the system will automatically generate an event report and notify the on-duty operation and maintenance personnel via SMS or APP notification. The generated event report content includes the error code and number, the name of the sensor involved, and the site information involved in the error; The output module uses the initial state of the equipment before adjustment and related environmental data as the starting point of reasoning after each data aggregation decision module and the site coordination execution module perform site equipment parameter adjustment optimization. It records the rules and algorithms based on which the data aggregation decision module formulates the optimization strategy, and each step in the process of executing the optimization strategy by the site coordination execution module, including the order of parameter adjustment, the adjustment range, and the feedback data of the equipment after each adjustment, and outputs a complete and detailed text description of the optimization strategy reasoning process data.

8. An agent-based unmanned intelligent operation and maintenance centralized control method for oilfield stations, characterized in that: The method adopts the agent-based unmanned intelligent operation and maintenance centralized control system for oilfield stations according to any one of claims 1 to 7, comprising the following steps: S1. By deploying pressure, temperature, flow, liquid level, and water content sensors at oilfield inter-well stations, transfer stations, and joint stations, real-time data on crude oil transportation equipment at these stations is collected. The data is received from subordinate inter-well stations, transfer stations, and joint stations in real time. Based on the data changes from real-time sensors at these stations, the LMM large model is used to generate cross-site instant optimization and adjustment strategies for transfer stations and joint stations, as well as target optimization and adjustment strategies. S2: Based on the real-time optimization and target optimization strategies for each site, and in combination with the data optimization models deployed at each site for deep optimization and fine-grained control of specific equipment and functions, the corresponding adjustment strategies are implemented for each site. Furthermore, based on data changes during the execution of the small models at different sites, the large-scale model optimization and adjustment strategies are further fine-tuned based on real-time monitoring data. S3. When the model reports an error, it automatically generates an event report and notifies the on-duty operation and maintenance personnel via SMS or APP notification. The generated event report includes the error code and number, the name of the sensor involved, and the site information involved in the error. S4. After each site equipment parameter adjustment and optimization, use the initial state of the equipment before adjustment and related environmental data as the starting point for reasoning, record the rules and algorithms based on which the optimization strategy is formulated, and record each step in the site's execution of the optimization strategy, including the order of parameter adjustment, the magnitude of adjustment, and the feedback data from the equipment after each adjustment. Output a complete text description of the optimization strategy reasoning process data.

Images