Multi-objective intelligent control method and system for product oil pipeline based on stage operation

Abstract

The application relates to the technical field of pipeline control, and discloses a product oil pipeline multi-target intelligent control method and system based on stage operation, which comprises the following steps: establishing a stage object to parse a multi-stage oil transportation plan into a structured task, obtaining real-time operation parameters and plan parameters, combining the structured task to generate a stage operation state, calculating reference throughputs of all participating stations under the conditions of equipment boundaries and hydraulic constraints, generating a reference setting, processing the stage operation state and the reference setting based on an expert model library and deep reinforcement learning SAC to obtain a control strategy, performing closed-loop adjustment on a cutting timing and an advance amount in a mixed oil cutting process, correcting the throughput distribution of a subsequent stage according to a stage execution deviation after the stage is completed, and uploading stage operation data to the expert model library and the deep reinforcement learning SAC for updating. The application realizes whole-process closed-loop optimization of multi-target operation of product oil pipelines, and improves safety and intelligence.

Description

Technical Field

[0001] This invention relates to the field of pipeline control technology, and more specifically, to a multi-objective intelligent control method and system for refined oil pipelines based on phased operation. Background Technology

[0002] Refined oil pipelines are a crucial infrastructure for national energy transportation, and their operational efficiency and safety directly impact the stability of energy supply. With the expansion of pipeline scale and the increasing complexity of oil transportation plans, the industry is gradually shifting from manual inspection to automated control. For example, patent CN103216729A proposes an intelligent control system for pipeline transportation. By adjusting the external pump through pressure sensors, level gauges, and frequency converters, it achieves an initial transition from manual inspection to automatic control, reducing labor intensity and improving operational stability to some extent. Simultaneously, with the application of PCS (Process Control System) systems and black-screen control modes, the industry has placed higher demands on one-click execution, closed-loop control throughout the entire process, and intelligent scheduling.

[0003] However, existing pipeline transportation control technologies still have several shortcomings. Their functions are mostly limited to single-point parameter control, lacking global analysis and dynamic closed-loop optimization of multi-stage oil transportation plans, leading to station asynchrony and low plan execution efficiency. Mixed oil cutting strategies still rely on manual experience, making it difficult to achieve precise closed-loop adjustment of interface detection and cutting timing, resulting in significant mixed oil losses. Current systems generally lack a comprehensive balance between multiple objectives such as energy consumption, mixed oil losses, and synchronous arrival, resulting in insufficient intelligence. Furthermore, traditional architectures have poor compatibility, making it difficult to introduce algorithms such as deep learning and reinforcement learning, thus hindering the development of intelligent and automated pipelines.

[0004] Therefore, it is necessary to design a multi-objective intelligent control method and system for refined oil pipelines based on phased operation to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes a multi-objective intelligent control method and system for refined oil pipelines based on phased operation, aiming to solve the problem of high safety risks caused by high dependence on manual operation in the current technology.

[0006] In one aspect, this invention proposes a multi-objective intelligent control method for refined oil pipelines based on phased operation, comprising:

[0007] Establish phase objects and parse the multi-stage oil transportation plan into structured tasks. The structured tasks include phase target volume, allowable deviation, participating stations, oil mixing and cutting strategy and phase switching conditions.

[0008] Real-time operating parameters and planned parameters are obtained through data fusion. The real-time operating parameters include pressure, flow rate, valve position, pump frequency and oil interface information, and are combined with the operating status of the structured task generation stage.

[0009] Under the conditions of equipment boundary and hydraulic constraints, based on the stage operation status and structured tasks, the reference flow of each participating station is calculated through flow balance to meet the stage target flow, minimize oil mixing loss and energy consumption constraints, and generate reference settings, which include reference pump frequency and reference valve position settings.

[0010] The stage operation state and reference settings are processed based on an expert model library and a deep reinforcement learning SAC (Soft Actor-Critic algorithm). The expert model library generates rule controls for pump allocation, flow regulation, oil mixing and abnormal handling, and the deep reinforcement learning SAC generates dynamic optimization continuous adjustment quantities. The control strategy is obtained based on the rule controls and continuous adjustment quantities.

[0011] During the oil-mixed cutting process, the cutting timing and advance are adjusted in a closed loop by combining the oil interface detection information. When the deviation exceeds the threshold, the cutting parameters are corrected.

[0012] After a phase is completed, the input allocation for subsequent phases is corrected based on the phase execution deviation, and the phase running data is uploaded to the expert model library and the deep reinforcement learning SAC for updating.

[0013] Furthermore, when establishing phase objects and parsing multi-phase oil transportation plans into structured tasks, this includes:

[0014] Based on the equipment boundary and hydraulic constraints, the consistency of the stage target quantity and the participating stations is verified. If it is not feasible, the stage is decomposed or merged according to priority and the stage target quantity is redistributed.

[0015] Based on the quantile statistics of tracking errors in historical periods and the risk level, an allowable deviation is generated, wherein the allowable deviation is 0.5% to 1.5%.

[0016] The cutting timing and lead time of the mixed oil cutting strategy are determined based on interface velocity estimation and density difference threshold, and the action sequence of the cutting valve is given.

[0017] The phase switching conditions are jointly triggered by the cumulative amount of the phase reaching the phase target amount minus the allowable deviation, the interface reaching the target position, and the safety status being met, and a hysteresis window is set.

[0018] Furthermore, real-time operating parameters and planned parameters are obtained through data fusion, and when combined with the operating status of the structured task generation stage, this includes:

[0019] Real-time operating parameters and planning parameters are collected by the RTU (Remote Terminal Unit) through OPC UA (Open Platform Communications Unified Architecture), and time synchronization is performed using PTP (Precision Time Protocol) to ensure that the clock deviation is no more than 100 milliseconds;

[0020] Outlier removal and 5-11 point sliding median noise reduction were performed on the pressure, flow rate, valve position, pump frequency and oil interface information to achieve unit and dimension unification.

[0021] Spline interpolation is performed on the missing measurement interval of no more than 60 seconds; zero-point and span calibration is performed on the sensor; multi-source weighted fusion is performed on the oil interface information and the interface estimate is output;

[0022] The phase target quantity, allowable deviation, participating stations and phase switching conditions are mapped to a unified time axis, and the phase remaining quantity, target deviation and remaining time estimate for inter-station synchronous arrival are calculated.

[0023] The stage operating status is generated using Kalman filtering with a period of 1 to 5 seconds. The stage operating status includes the pressure vector inside and outside the station, the inlet and outlet flow rate, the valve position, the pump frequency, the oil interface estimation, the stage remaining quantity, the target deviation, the safety margin, and the data quality label.

[0024] Furthermore, under the conditions of equipment boundaries and hydraulic constraints, based on the aforementioned stage operating status and structured tasks, the reference throughput of each participating station is calculated through throughput balancing to generate the reference setting, including:

[0025] Using the aforementioned stage operation status and structured tasks as input, the minimum remaining arrival time deviation between stations is taken as the synchronization index, and oil mixing loss and energy consumption are taken as parallel objectives.

[0026] Under the constraints of equipment boundaries and hydraulic constraints, a quadratic programming approach is used to solve for the reference throughput of each participating station. The equipment boundaries include the range and rate of change of pump frequency and the range and rate of change of valve position. The hydraulic constraints include the upper limit of pipe section pressure, the minimum required net positive suction head, and the upper limit of pipe velocity.

[0027] Based on the pump characteristic curve and the valve flow coefficient, the reference output is mapped to the reference pump frequency and the reference valve position setting to form the reference setting.

[0028] Furthermore, the expert model library stores pump allocation models, flow regulation models, mixed oil cutting models, and abnormal handling models using decision trees or rule sets. Taking the stage operation status and reference settings as input, it controls the output according to preset trigger rules. The rule control includes pump frequency setting increment, valve position setting increment, and action sequence, and performs boundary checks on stage switching conditions and allowable deviations. When rules conflict, the decision is made in the order of safety priority > mixed oil cutting > flow regulation > pump allocation.

[0029] Furthermore, the deep reinforcement learning SAC takes the stage operation state and reference settings as input; the action is a continuous pump frequency increment and valve position increment, and is constrained by step size and range; the reward is mainly based on minimizing the error of tracking the reference settings, while adding oil mixing loss, energy consumption penalty and limit violation penalty; it is trained offline in a simulation environment and outputs a continuous adjustment amount.

[0030] Furthermore, when acquiring control strategies, this includes:

[0031] First, the continuous adjustment amount is projected into the feasible region according to the safety constraints controlled by the rules, and then fused with the rule control according to the preset weights to form the control strategy;

[0032] When there is a risk of overpressure, cavitation, or interface misalignment in short-term predictions based on the stage of operation, rule-based control is used instead of continuous adjustment to ensure that the control strategy meets the requirements of synchronous arrival, oil mixing loss suppression, and energy consumption constraints.

[0033] Furthermore, by combining oil interface detection information, a closed-loop adjustment is performed on the cutting timing and lead time. When the deviation exceeds a threshold, the cutting parameters are corrected, including:

[0034] Using the stage operation status and the oil interface estimate obtained by data fusion as input, a cutting sequence and advance amount are generated, wherein the advance amount is 30 to 90 seconds.

[0035] The valve cutting action sequence is invoked according to the reference settings. The action sequence includes opening, holding, and closing, and the valve position change rate and minimum holding time are limited to 2 to 5 seconds.

[0036] The short-window prediction corrector reads the oil interface detection information and predicts the interface arrival time at a cycle of 1 to 2 seconds, and corrects the cutting sequence and advance amount.

[0037] After the cutting is completed, the deviation is determined based on the interface positioning deviation or the oil mixing trail index. When the deviation exceeds the cutting deviation threshold, the cutting parameters are corrected. The cutting parameters include the lead amount, valve opening degree and holding time. The correction result is written into the oil mixing cutting strategy of the next stage.

[0038] Furthermore, after a phase is completed, when correcting the input allocation based on the phase execution deviation and uploading the running data to the expert model library and the deep reinforcement learning SAC for updates, the process includes:

[0039] Summarize the stage operation status, reference settings, control strategies and stage cumulative quantities, and calculate the stage execution deviation, which includes target deviation, synchronous arrival deviation, oil mixing and cutting deviation and energy consumption deviation.

[0040] Based on the stage execution deviation, a correction factor is generated to proportionally limit the allocation of stage target quantities and reference settings in subsequent stages, while maintaining the consistency of stage switching conditions and allowable deviations.

[0041] The operational data, including stage operation status, reference settings, rule control, continuous adjustment amount, control strategy and stage execution deviation, is uploaded to the expert model library and the deep reinforcement learning SAC. The expert model library performs amplitude-limited updates on rule thresholds and action timing, and the deep reinforcement learning SAC uses the operational data as experience replay samples for incremental training.

[0042] Compared with existing technologies, the advantages of this invention are as follows: By introducing stage objects, the complex multi-stage oil transportation plan is analyzed into a structured task that includes target volume, allowable deviation, participating stations, oil mixing and cutting strategies, and stage switching conditions. This structured task serves as the core driver to achieve closed-loop control throughout the entire process, overcoming the limitation of existing technologies that can only adjust at single points. Through data fusion, the stage operation status is constructed, and flow balance optimization is performed under equipment boundaries and hydraulic constraints. This not only achieves synchronous arrival between stations but also minimizes oil mixing losses and energy consumption constraints, improving global optimization capabilities. Utilizing the collaborative processing mechanism of an expert model library and deep reinforcement learning SAC, the advantages of empirical rules and adaptive optimization are combined, ensuring safety and robustness, and providing continuous learning and dynamic optimization capabilities. In the oil mixing and cutting stage, closed-loop adjustment of timing and lead time is achieved by combining oil interface detection, reducing oil mixing losses and cutting deviations. After the stage is completed, deviation feedback is executed to correct the flow allocation and update the expert model library and SAC strategy, forming a continuously evolving intelligent control system. This achieves closed-loop optimization of the entire process of multi-objective operation of refined oil pipelines, combining safety, economy, and intelligence.

[0043] On the other hand, this application also provides a multi-objective intelligent control system for refined oil pipelines based on phased operation, used to apply the above-mentioned multi-objective intelligent control method for refined oil pipelines based on phased operation, including:

[0044] The data acquisition unit is configured to establish phase objects and parse the multi-phase oil transportation plan into structured tasks. The structured tasks include phase target quantities, allowable deviations, participating stations, oil mixing and cutting strategies, and phase switching conditions.

[0045] The acquisition unit is also configured to acquire real-time operating parameters and planned parameters through data fusion. The real-time operating parameters include pressure, flow rate, valve position, pump frequency and oil interface information, and are combined with the operating status of the structured task generation stage.

[0046] The first processing unit is configured to, under the conditions of equipment boundary and hydraulic constraints, calculate the reference throughput of each participating station through throughput balance based on the stage operation status and structured tasks, so as to meet the stage target throughput of synchronous arrival between stations, minimize oil mixing loss and energy consumption constraints, and generate reference settings, the reference settings including reference pump frequency and reference valve position settings.

[0047] The second processing unit is configured to process the stage operation status and reference settings based on the expert model library and deep reinforcement learning SAC. The expert model library generates rule control for pump allocation, flow regulation, oil mixing and abnormal handling, and the deep reinforcement learning SAC generates dynamic optimization continuous adjustment amount to finally obtain the control strategy.

[0048] The adjustment unit is configured to perform closed-loop adjustment of the cutting timing and advance amount in combination with oil interface detection information during the oil mixing cutting process, and to correct the cutting parameters when the deviation exceeds the threshold.

[0049] The update unit is configured to perform deviation correction of input allocation according to the stage after the stage is completed, and upload the running data to the expert model library and deep reinforcement learning SAC for updating.

[0050] It is understandable that the above-mentioned multi-objective intelligent control method and system for refined oil pipelines based on phased operation have the same beneficial effects, and will not be elaborated further here. Attached Figure Description

[0051] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0052] Figure 1 A flowchart of a multi-objective intelligent control method for refined oil pipelines based on phased operation, provided in an embodiment of the present invention;

[0053] Figure 2 This is a functional block diagram of a multi-objective intelligent control system for refined oil pipelines based on phased operation, provided in an embodiment of the present invention. Detailed Implementation

[0054] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0055] In traditional refined oil pipeline control systems, single-point parameter control modes cannot effectively analyze the global task structure of multi-stage oil transportation plans, resulting in a lack of dynamic coordination between stage target quantities, participating stations, and oil mixing and cutting strategies. The system relies on manual experience to set the oil mixing and cutting sequence and lead time; the interface detection information and cutting action sequence do not form a closed-loop adjustment mechanism, leading to accumulated tracking deviations at the oil mixing interface. Furthermore, the multi-objective optimization process lacks a balance mechanism between inter-station synchronous arrival and energy consumption constraints; the capacity allocation calculation does not consider the dynamic coupling effect of equipment boundaries and hydraulic constraints, causing conflicts between pump frequency and valve position adjustments.

[0056] For example, in long-distance refined oil pipeline systems, multiple pumping stations need to coordinate to execute oil transportation plans that include oil product switching. The existing control architecture treats each station as an independent unit, failing to parse the stage target quantities into structured task parameters, resulting in the allocation of throughput to each station not being adjusted synchronously according to the remaining time. During the mixed oil cutting process, the manually preset cutting lead time is not combined with real-time interface speed estimation, leading to a lag between the cutting valve action sequence and the arrival time of the oil interface. The throughput balance calculation does not incorporate joint constraints of pump frequency range, valve position change rate, and pipeline section pressure upper limit, causing some pumping stations to operate under overpressure while other pumping stations have insufficient suction pressure.

[0057] If the above issues are not resolved, multi-stage oil transportation plans will fail to meet the stage switching conditions due to asynchrony between stations, leading to batch execution interruptions or plan rollbacks. The continuous accumulation of mixing and cutting deviations will increase the length of the mixing section, causing oil quality contamination and economic losses. Conflicts between pump frequency and valve position adjustments may trigger protective shutdowns of equipment, reducing pipeline operational continuity. When the dynamic coupling effect of hydraulic constraints is not suppressed, pressure fluctuations within the pipeline will exacerbate cavitation risks, threatening the safe operation of the pipeline.

[0058] For this, please refer to Figure 1 As shown, this application proposes a multi-objective intelligent control method for refined oil pipelines based on phased operation, including:

[0059] S100: Establish phase objects and parse the multi-phase oil transportation plan into structured tasks. The structured tasks include phase target volume, allowable deviation, participating stations, oil mixing and cutting strategies, and phase switching conditions.

[0060] S200: Real-time operating parameters and planned parameters are obtained through data fusion. Real-time operating parameters include pressure, flow rate, valve position, pump frequency and oil interface information, and are combined with the operating status of the stage generated by structured task generation.

[0061] S300: Under the conditions of equipment boundary and hydraulic constraints, based on the stage operation status and structured tasks, the reference flow of each participating station is calculated through flow balance to ensure that the station synchronously reaches the stage target flow, minimizes oil mixing loss and energy consumption constraints, and generates reference settings, including reference pump frequency and reference valve position settings.

[0062] S400: Based on the expert model library and deep reinforcement learning SAC, the stage operation status and reference settings are processed. The expert model library generates rule control for pump allocation, flow regulation, oil mixing and abnormal handling, while the deep reinforcement learning SAC generates dynamic optimization continuous adjustment quantity. The control strategy is obtained based on the rule control and continuous adjustment quantity.

[0063] S500: During the oil-mixed cutting process, the cutting timing and advance amount are adjusted in a closed loop based on the oil interface detection information. When the deviation exceeds the threshold, the cutting parameters are corrected.

[0064] S600: After a phase is completed, the input allocation for subsequent phases is corrected based on the phase execution deviation, and the phase running data is uploaded to the expert model library and the deep reinforcement learning SAC for updating.

[0065] Specifically, the "stage object" refers to transforming a multi-stage oil transportation plan into a structured task that includes stage target quantities, allowable deviations, participating stations, oil mixing strategies, and stage switching conditions. This can be implemented using analytical algorithms or task decomposition models to achieve standardization and executability of the multi-stage plan. The "stage operation status" refers to dynamic operation indicators generated by fusing real-time and planned parameters. This is achieved using data cleaning, multi-source information fusion, and Kalman filtering techniques to reflect the current stage's execution progress and safety margin. The "transportation balance calculation" refers to solving for the reference transport volume of each station under equipment boundaries and hydraulic constraints. This is achieved using a quadratic programming algorithm combined with pump and valve characteristic curves to ensure that stations synchronously reach the stage target quantity while satisfying oil mixing losses and energy consumption constraints. The "expert model library and deep reinforcement learning SAC collaborative processing" refers to generating control strategies through the fusion of rule-based control and continuous adjustment. This involves using decision trees to store pump allocation and flow regulation rules, combined with reinforcement learning simulation training to generate dynamic optimization quantities, achieving multi-objective dynamic adjustment under safety constraints. Closed-loop control of cutting timing and lead time refers to dynamically adjusting cutting parameters based on oil interface detection information. This is achieved using a short-window predictive corrector and a deviation threshold determination mechanism to reduce oil mixing cutting deviation. Stage execution deviation correction refers to adjusting the output allocation of subsequent stages based on the actual data after the completion of each stage. This is achieved using a proportional amplitude limiting correction algorithm and an experience playback mechanism to improve the robustness of multi-stage plan execution.

[0066] This application constructs a multi-stage intelligent control framework. By generating dynamic operating states through structured task parsing and data fusion, and combining output balance calculation with expert model library reinforcement learning for collaborative optimization, it achieves multi-objective dynamic closed-loop control of inter-station synchronous arrival, oil mixing loss suppression, and energy consumption constraints. This solves the problems of lack of global optimization, reliance on manual experience for oil mixing cutting, and difficulty in balancing multiple objectives in the existing technology.

[0067] The working process and principle of this application are as follows: First, a phase object is established, and the multi-stage oil transportation plan is parsed into structured tasks, including the phase target quantity, allowable deviation, participating stations, oil mixing cut-off strategy, and phase switching conditions. Real-time operating parameters and planned parameters, including pressure, flow rate, valve position, pump frequency, and oil interface information, are obtained through data fusion, and the phase operating status is generated in combination with the structured tasks. Under the conditions of equipment boundary and hydraulic constraints, based on the phase operating status and structured tasks, the reference flow rate of each participating station is calculated through flow balance to ensure that it meets the requirements of synchronous arrival of the phase target quantity between stations, minimization of oil mixing loss, and energy consumption constraints, and reference settings are generated, including reference pump frequency and reference valve position settings. The phase operating status and reference settings are processed based on an expert model library and deep reinforcement learning SAC. The expert model library generates rule control for pump allocation, flow regulation, oil mixing cut-off, and abnormal handling, and the deep reinforcement learning SAC generates dynamically optimized continuous adjustment quantities. The control strategy is obtained based on the rule control and continuous adjustment quantities. During the oil mixing cut-off process, the cutting sequence and advance amount are adjusted in a closed loop in combination with oil interface detection information. When the deviation exceeds the threshold, the cutting parameters are corrected. After a phase is completed, the input allocation for subsequent phases is adjusted based on the phase execution deviation, and the phase execution data is uploaded to the expert model library and the deep reinforcement learning SAC for updates.

[0068] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0069] When establishing phase objects, the multi-phase oil transportation plan is parsed into structured tasks. These structured tasks include the phase target quantity, allowable deviation, participating stations, blending and cutting strategies, and phase switching conditions. The phase target quantity is the amount of oil that each participating station needs to transport in the current phase. The allowable deviation is the permissible error range for the phase target quantity. Participating stations are the list of stations participating in oil transportation in the current phase. The blending and cutting strategy includes the cutting sequence and lead time. The phase switching conditions are the combination of conditions that trigger the start of the next phase.

[0070] Real-time and planned operating parameters are obtained through data fusion. Real-time operating parameters include pressure, flow rate, valve position, pump frequency, and oil interface information. Planned parameters include preset parameters in the oil transportation plan. Data fusion employs a multi-source information fusion algorithm to comprehensively process data from different sources. This is combined with the operational status of the structured task generation stage, including the current stage's execution progress and the real-time values ​​of each parameter.

[0071] Under equipment boundary and hydraulic constraints, and based on the stage operation status and structured tasks, reference throughput for each participating station is calculated through throughput balancing. The throughput balancing calculation considers three objectives: synchronous arrival of the stage target throughput between stations, minimization of mixing losses, and energy consumption constraints. A multi-objective optimization algorithm is used to solve for the optimal reference throughput. Reference settings are generated, including reference pump frequency and reference valve position settings.

[0072] The system processes the stage operation status and reference settings based on an expert model library and a deep reinforcement learning-based System-Action Control (SAC). The expert model library stores pre-set rule models used to generate rule-based controls for pump allocation, flow regulation, oil mixing, and anomaly handling. The deep reinforcement learning-based SAC generates dynamically optimized continuous adjustment quantities through reinforcement learning algorithms. Combining rule-based control with these continuous adjustment quantities yields the final control strategy.

[0073] During the oil mixing process, the cutting timing and lead time are adjusted in a closed loop based on oil interface detection information. An interface detection instrument monitors the oil interface position in real time. When the detected interface deviation exceeds a preset threshold, the cutting parameters, including the cutting timing and lead time, are dynamically corrected.

[0074] After a phase is completed, the input allocation for subsequent phases is adjusted based on the phase execution deviation. The deviation between the actual execution of the current phase and the planned performance is calculated, and the input allocation for subsequent phases is adjusted accordingly. Phase execution data is uploaded to the expert model library and the Deep Reinforcement Learning (SAC) for model updates and algorithm optimization.

[0075] Through the above scheme, this application achieves structured analysis and dynamic closed-loop optimization of multi-stage oil transportation plans, improving plan execution efficiency. A precise closed-loop adjustment mechanism for oil mixing and cutting is established, reducing oil mixing losses. Through multi-objective comprehensive optimization, a balance is achieved between multiple objectives such as energy consumption, oil mixing losses, and synchronous arrival. The introduction of deep learning and reinforcement learning algorithms enhances the intelligence level of pipeline control.

[0076] In some of the schemes mentioned above in this application, the process of establishing phase objects and parsing multi-stage oil transportation plans into structured tasks may lead to the plan being infeasible due to the mismatch between the phase target quantity and equipment capacity. At the same time, the timing and lead time of the oil mixing and cutting strategy lack a dynamic adjustment mechanism, and the single phase switching condition may cause misjudgment or delay.

[0077] This application further proposes that when establishing phase objects, consistency verification of phase target quantities and participating stations should be performed based on equipment boundaries and hydraulic constraints. If this is not feasible, phases should be decomposed or merged according to priority and phase target quantities should be redistributed. Allowable deviations should be generated based on quantile statistics of historical phase tracking errors and risk levels, with an allowable deviation of 0.5% to 1.5%. The cutting sequence and advance amount of the mixed oil cutting strategy should be determined based on interface velocity estimation and density difference threshold, and the cutting valve action sequence should be given. The phase switching conditions should be jointly triggered by the phase cumulative quantity reaching the phase target quantity minus the allowable deviation, the interface reaching the target position, and the safety status being met, and a hysteresis window should be set.

[0078] The equipment boundaries and hydraulic constraints include parameters such as pump frequency range, valve position range, and upper limit of pipeline pressure. Consistency verification is achieved by comparing the stage target quantity with the equipment's maximum delivery capacity; if the limit is exceeded, the stage is decomposed or the target quantity is adjusted. The quantile statistics of historical stage tracking errors use the 90th percentile as the basis for risk level classification, with higher risk levels corresponding to smaller allowable deviations. Interface velocity estimation is calculated through multi-source sensor data fusion, and the density difference threshold is set according to the oil type; when the density difference exceeds the threshold, a cutting action is triggered. The joint triggering of stage switching conditions uses a logical AND relationship, and the hysteresis window is set to 0.1% to 0.3% of the stage target quantity.

[0079] Specifically, during the establishment of the phase object, the equipment boundary parameters and hydraulic constraints are first extracted. For example, the upper limit of pump frequency at a certain station is 50Hz, and the valve opening range is 20% to 80%. If the theoretical throughput corresponding to the phase target quantity requires a pump frequency of 55Hz, it is determined to be infeasible. In this case, the original phase is decomposed into two sub-phases according to priority, with the target quantities allocated to 60% and 40% of the original value, respectively. After quantile statistics of historical tracking error data, if the 90th percentile is 1.2%, the allowable deviation is set to 1.0%. The interface velocity is calculated to be 2.5m / s through the fusion of ultrasonic and density meter data. Combined with the density difference threshold of 0.5kg / m³, the cutting advance is determined to be 45 seconds. In the phase switching conditions, the phase switching is triggered when the cumulative phase quantity reaches 99.5% of the target quantity, the interface reaches the inlet of the target storage tank, and the pressure fluctuation is less than 0.2MPa. The hysteresis window is set to 0.2% of the target quantity. Through the above steps, the feasibility of the phase plan is ensured, and the dynamic optimization of mixed oil cutting and the accurate judgment of switching conditions are achieved.

[0080] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0081] When establishing phase objects, the multi-phase oil transportation plan is parsed into structured tasks. First, the consistency between the phase target quantity and participating stations is checked based on equipment boundaries and hydraulic constraints. If infeasibility is found, phases are decomposed or merged according to priority, and the phase target quantity is reallocated. For example, when the target quantity of a certain phase exceeds the equipment capacity, the phase can be decomposed into two sub-phases, or merged with adjacent phases and the target quantity is reallocated.

[0082] Furthermore, permissible deviations are generated based on quantile statistics of tracking errors from historical periods and risk levels. Specifically, the 90th percentile can be used as the permissible deviation benchmark, and adjusted according to different risk levels. For high-risk levels, the permissible deviation is set at 0.5%, for medium-risk at 1%, and for low-risk at 1.5%.

[0083] Therefore, the cutting timing and advance amount of the mixed oil cutting strategy are determined based on the interface velocity estimation and density difference threshold. For example, when the interface velocity estimation is 1.2 m / s and the density difference threshold is 5 kg / m³, the cutting advance amount can be set to 60 seconds. The cutting valve action sequence includes three steps: opening, holding, and closing.

[0084] Finally, the stage switching conditions are jointly triggered by the cumulative amount of each stage reaching the stage target amount minus the allowable deviation, the interface reaching the target position, and the safety status being met. As a preferred implementation, a 5-second hysteresis window can be set to avoid frequent switching.

[0085] Through the above technical solutions, this application achieves structured analysis and dynamic optimization of multi-stage oil transportation plans. Consistency checks and target quantity reallocation improve plan executability. The allowable deviation generated based on historical data is more reasonable, contributing to improved tracking accuracy. Optimization of the oil mixing and cutting strategy reduces oil mixing losses. Jointly triggered stage switching conditions ensure safe and stable operation. This enhances the execution efficiency and accuracy of multi-stage oil transportation.

[0086] In some of the solutions described above in this application, when obtaining real-time operating parameters and generating planned parameters through data fusion, there are problems such as asynchronous data acquisition, sensor noise interference, and oil interface detection errors, which lead to insufficient reliability of the stage operating status and affect the accuracy of subsequent transport balance calculations.

[0087] This application further proposes a method for obtaining real-time operating parameters and planned parameters through data fusion, and generating stage operating status by combining structured tasks. This includes: real-time operating parameters and planned parameters are collected by the RTU through OPC UA; PTP time synchronization is used to ensure that the clock deviation is no more than 100 milliseconds; outlier removal and 5-11 point sliding median denoising are performed on pressure, flow, valve position, pump frequency and oil interface information to achieve unit and dimension unification; spline interpolation is performed on the missing measurement interval of no more than 60 seconds; zero point and span calibration of sensors is performed; multi-source weighted fusion of oil interface information is performed and the interface estimate is output; the stage target quantity, allowable deviation, participating stations and stage switching conditions are mapped to a unified time axis; the stage remaining quantity, target deviation and remaining time estimate for inter-station synchronous arrival are calculated; and stage operating status is generated by Kalman filtering with a period of 1-5 seconds. The stage operating status includes internal and external pressure vectors, inlet and outlet flow rates, valve positions, pump frequencies, oil interface estimates, stage remaining quantity, target deviation, safety margin and data quality markers.

[0088] The RTU achieves multi-source heterogeneous data acquisition via the OPC UA protocol, while the PTP time synchronization protocol controls clock deviation to within 100 milliseconds, eliminating time drift of distributed nodes. Sliding median denoising uses odd-point windows to filter pressure and flow data, effectively suppressing impulse noise. Spline interpolation reconstructs continuous data streams within a 60-second missing measurement interval, avoiding interruptions in state calculations due to communication interruptions. Multi-source weighted fusion assigns weight coefficients based on the historical error rates of each oil interface sensor, reducing the risk of single sensor failure. Kalman filtering fuses structured task parameters and real-time data at a fixed period to generate a state vector containing data quality markers.

[0089] Specifically, the RTU collects pressure, flow, valve position, pump frequency, and oil interface data via the OPC UA protocol, and synchronizes clocks at each station via the PTP protocol to ensure time alignment of cross-station data. Pressure data undergoes an 11-point sliding median filter to eliminate abnormal spikes, while flow data is smoothed using a 5-point window. When a sensor experiences a short-term failure within 60 seconds, cubic spline interpolation is used to reconstruct the data curve for the missing interval. Oil interface detection data integrates multi-source measurements from the density meter and ultrasonic interface meter using a weighted fusion algorithm, outputting an interface estimate with an error rate of less than 0.3%. After the structured task parameters are mapped to a unified time axis, they are combined with real-time acquired data to calculate the remaining quantity, target deviation, and remaining time estimate. A Kalman filter fuses the above parameters at 1-second intervals to generate a state vector containing pressure vector, flow rate, valve position, pump frequency, and oil interface estimate, while simultaneously outputting a quality marker bit representing data integrity.

[0090] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0091] Real-time operating and planned parameters are acquired from the RTU via the OPC UA protocol. PTP time synchronization technology is used to control clock deviation within 100 milliseconds. Outlier removal is performed on the acquired pressure, flow, valve position, pump frequency, and oil interface information, and noise reduction is achieved using the 7-point moving median method. After unifying units and dimensions, cubic spline interpolation is performed on missing data intervals not exceeding 60 seconds. Zero-point and span calibration of the sensors is performed to improve data accuracy. A multi-source weighted fusion algorithm is used for the oil interface information to output an estimated interface position.

[0092] The phase target quantity, allowable deviation, participating stations, and phase switching conditions are mapped onto a unified time axis. The phase remaining quantity, target deviation, and estimated remaining time for inter-station synchronous arrival are calculated. The phase operating status is generated using a Kalman filter algorithm with a 3-second cycle. The phase operating status includes internal and external pressure vectors, inlet and outlet flow rates, valve positions, pump frequencies, oil interface estimation, phase remaining quantity, target deviation, safety margin, and data quality markers.

[0093] Through the above technical solutions, this application achieves high-precision acquisition and processing of real-time and planned operating parameters. By employing multi-source data fusion and filtering algorithms, the accuracy of oil interface position estimation is improved. Using a unified time axis mapping and Kalman filtering, stage-specific operating status information comprehensively reflecting the pipeline's operating state is generated. This provides a high-quality data foundation for subsequent intelligent control decisions, effectively improving the reliability and accuracy of multi-objective intelligent control of refined oil pipelines.

[0094] In some of the solutions described above in this application, after generating the operational status through data fusion, a reference setting needs to be generated based on this status and the structured task. However, existing methods struggle to simultaneously meet the multi-objective optimization requirements of synchronous arrival between stations, minimizing oil mixing losses, and energy consumption constraints when calculating the reference throughput, leading to low oil transfer efficiency or unstable equipment operation.

[0095] This application further proposes to use the stage operation status and structured tasks as inputs, the minimum remaining arrival time deviation between stations as the synchronization index, and the mixing loss and energy consumption as parallel objectives. Under the constraints of equipment boundaries and hydraulic constraints, a quadratic programming approach is used to solve for the reference throughput of each participating station. The equipment boundaries include the pump frequency range and rate of change, and the valve position range and rate of change. The hydraulic constraints include the upper limit of pipe section pressure, the minimum required net positive suction head, and the upper limit of pipe flow velocity. Based on the pump characteristic curve and valve flow coefficient, the reference throughput is mapped to the reference pump frequency and reference valve position settings to form the reference settings.

[0096] In the quadratic programming solution process, synchronization indicators, oil mixing losses, and energy consumption are treated as different components of the objective function, and a multi-objective balance is achieved through weighting coefficients. Equipment boundaries are constrained by inequality constraints to limit the feasible range of pump frequency and valve position, while hydraulic constraints ensure pipeline operational safety through pressure and flow velocity limitations. Pump characteristic curves and valve flow coefficients transform theoretical throughput into executable equipment parameters, such as setting the pump frequency range to 25–50 Hz and the valve position change rate not exceeding 5% per second.

[0097] Specifically, the pressure vectors inside and outside the station during the phased operation, the inflow and outflow flow data, and the phased objective quantities from the structured task are all input into the quadratic programming model. The objective function consists of three parts: the square of the remaining arrival time deviation between stations, the linear term for oil mixing loss, and the quadratic term for energy consumption. Constraints include a pump frequency change rate not exceeding 2 Hz per second, valve opening limited to 10%–90%, and pipeline pressure not exceeding 8 MPa. After solving, the reference throughput for each station is obtained; for example, the reference throughput for a certain station is 1200 m³ / h. According to the pump characteristic curve, this throughput corresponds to a pump frequency of 42 Hz, and according to the valve flow coefficient, it corresponds to a valve opening of 65%. The resulting reference settings satisfy both the multi-objective optimization requirements and the equipment operating boundaries and hydraulic safety constraints.

[0098] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0099] Using the phased operational status and structured tasks as inputs, and minimizing the remaining arrival time deviation between stations as the synchronization indicator, oil mixing losses and energy consumption are considered as parallel objectives. Under equipment boundaries and hydraulic constraints, quadratic programming is used to solve for the reference throughput of each participating station. Equipment boundaries include the pump frequency range and rate of change, and the valve position range and rate of change. Hydraulic constraints include the upper limit of pipe section pressure, the minimum required net positive suction head, and the upper limit of pipe velocity. Based on the pump characteristic curves and valve flow coefficients, the reference throughput is mapped to reference pump frequency and reference valve position settings to form reference settings.

[0100] Specifically, a quadratic programming model is first established, with the objective function being to minimize the weighted sum of inter-station residual arrival time deviation, mixing losses, and energy consumption. Constraints include equipment boundaries and hydraulic constraints. Specifically, the pump frequency range is 20–50 Hz, with a pump frequency change rate not exceeding 0.5 Hz / s; the valve position range is 0%–100%, with a valve position change rate not exceeding 2% / s; the upper limit of pipeline pressure is 10 MPa; the minimum required net positive suction head is 3 m; and the upper limit of flow velocity within the pipeline is 3 m / s.

[0101] Furthermore, the interior-point method is used to solve the quadratic programming problem to obtain the reference throughput for each participating station. Then, based on the pre-calibrated pump characteristic curves and valve flow coefficients, the reference throughput is converted into corresponding reference pump frequencies and reference valve position settings. For example, for a centrifugal pump, its characteristic curve can be expressed as H = a - bQ. 2 Where H is the head, Q is the flow rate, and a and b are fitting coefficients. This characteristic curve allows the reference output Q to be mapped to the corresponding reference pump frequency. Similarly, for a certain control valve, the relationship between its flow coefficient Cv and opening θ can be expressed as Cv = k * θ. n , where k and n are fitting coefficients. This relationship allows the reference input Q to be mapped to the corresponding reference valve position setting.

[0102] This generates a reference setting that includes a reference pump frequency and a reference valve position setting, which serves as a benchmark for subsequent control.

[0103] Through the above technical solution, this application achieves reference setting generation based on multi-objective optimization. By solving a quadratic programming problem, a balanced reference throughput is obtained while simultaneously considering multiple objectives such as inter-station synchronization, mixing losses, and energy consumption, all while satisfying equipment boundary and hydraulic constraints. This reference throughput is then mapped to specific equipment control parameters. Compared to single-objective optimization, this approach more comprehensively balances multiple factors, improving the overall efficiency and economy of pipeline operation. Furthermore, by considering equipment and hydraulic constraints, the feasibility and safety of the generated reference settings are ensured.

[0104] In some of the solutions described above in this application, rule control generated by the expert model library may lead to decision conflicts due to the parallel triggering of multiple models, resulting in unstable control strategies or reduced execution efficiency.

[0105] This application further proposes an expert model library that stores pump allocation models, flow regulation models, oil mixing and cutting models, and abnormal handling models using decision trees or rule sets. The model takes the stage operation status and reference settings as inputs and controls the output according to preset trigger rules. The rule control includes pump frequency setting increments, valve position setting increments, and action sequences. It also performs boundary checks on stage switching conditions and allowable deviations. When rules conflict, the decision is made in the order of safety priority > oil mixing and cutting > flow regulation > pump allocation.

[0106] The decision tree or rule set storage method supports independent operation and cross-triggering of multiple models. Preset triggering conditions are based on a combination of thresholds for stage remaining quantity, target deviation, and safety margin. The pump allocation model generates pump frequency setting increments based on pump group efficiency curves; the flow regulation model adjusts valve position setting increments based on the remaining time of inter-station synchronous arrival; the mixed oil cutting model generates valve action sequences through interface speed prediction; and the anomaly handling model triggers safety protection actions based on pressure deviation and data quality markers. A hierarchical arbitration mechanism is used to resolve rule conflicts. The anomaly handling model corresponding to safety priority has the highest decision weight, followed by the mixed oil cutting model, with the priority decreasing progressively for flow regulation and pump allocation models.

[0107] Specifically, after the stage operation status and reference settings are input into the expert model library, each model independently generates pump frequency increments, valve position increments, and action sequences according to preset rules. For example, when the pressure vector exceeds the upper limit of the pipe section, the abnormal handling model triggers safety protection rules, generating a valve closing or frequency reduction command. At the same time, the mixed oil cutting model may generate a valve opening command due to the interface reaching the prediction. In this case, the valve closing command of the abnormal handling model is executed first, suspending the mixed oil cutting action, in the order of safety priority > mixed oil cutting > flow regulation > pump allocation. When verifying the stage switching conditions, if the stage cumulative amount reaches the target amount minus the allowable deviation, the stage switching rule is triggered, generating a pump frequency zeroing and valve closing sequence. The boundary verification process limits the pump frequency change rate of the reference setting to prevent it from exceeding the equipment boundary. Through a hierarchical decision-making mechanism, it is ensured that the control strategy prioritizes safety constraints under complex operating conditions, while also taking into account the goals of mixed oil loss suppression and energy consumption optimization.

[0108] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0109] The expert model library stores pump allocation models, flow regulation models, mixed oil cutting models, and anomaly handling models using decision trees or rule sets. It takes the stage operating status and reference settings as input and controls output according to preset trigger rules. Rule control includes pump frequency setting increments, valve position setting increments, and action sequences, and performs boundary checks on stage switching conditions and allowable deviations. When rules conflict, the decision is made in the order of safety priority > mixed oil cutting > flow regulation > pump allocation.

[0110] Specifically, the pump matching model in the expert model library adopts a decision tree structure, selecting the optimal pump combination based on parameters such as current flow rate, pressure, and pump efficiency. The flow regulation model uses a rule set to provide corresponding valve opening adjustment amounts based on the magnitude and trend of flow deviation. The oil mixing and cutting model combines interface position and velocity information to generate the opening sequence and opening change curve of the cutting valve. The anomaly handling model includes multiple if-then rules, covering emergency response procedures for abnormal situations such as overpressure, underpressure, and leakage.

[0111] Furthermore, the output of rule-based control is incremental, such as increasing the pump frequency by 5Hz or decreasing the valve opening by 2%. The action sequence includes the timing arrangement of multiple discrete actions, such as first reducing the upstream pump frequency, then opening the cutting valve, and finally adjusting the downstream valve. Boundary checks ensure that the control actions do not cause the parameters to exceed the safe range.

[0112] Therefore, when multiple models are triggered simultaneously, decisions are made according to a predetermined priority order. For example, if flow regulation and oil mixing / cutting rules are triggered simultaneously, the control actions related to oil mixing / cutting are executed first. This hierarchical decision-making mechanism ensures stability and reliability under complex operating conditions.

[0113] Through the above technical solutions, this application achieves intelligent decision-making based on expert experience. The expert model library integrates operational experience and can quickly respond to various operating conditions. Rule control adopts an incremental approach, avoiding fluctuations caused by large-scale adjustments. The priority decision-making mechanism ensures the timely execution of critical operations, improving safety and reliability. Meanwhile, the boundary verification mechanism effectively prevents the risk of exceeding limits due to misoperation. Overall, this solution improves the intelligence and automation level of pipeline operation, reduces human intervention, and enhances operational efficiency and safety.

[0114] In some of the solutions mentioned above in this application, the expert model library generates discrete actions such as pump matching, flow regulation, oil mixing and abnormal handling through rule control. However, discrete actions are difficult to achieve continuous dynamic optimization, resulting in large errors in tracking the reference settings. Oil mixing losses and energy consumption are difficult to suppress effectively, and the continuous adjustment amount cannot be adaptively adjusted to adapt to complex working condition changes.

[0115] This application further proposes a deep reinforcement learning SAC that takes the stage operation state and reference settings as inputs, the actions as continuous pump frequency increments and valve position increments, and is subject to step size and range constraints. The reward is based on minimizing the error of tracking the reference settings. At the same time, it incorporates oil mixing loss, energy consumption penalty and limit violation penalty, and adopts offline training in a simulation environment to output continuous adjustment quantity.

[0116] The phased operation status includes internal and external pressure vectors, inlet and outlet flow rates, valve positions, pump frequencies, oil interface estimation, phase remaining quantity, target deviation, safety margin and data quality marking, reference settings including reference pump frequency and reference valve position settings, step size constraint of continuous pump frequency increment is ±0.5Hz, step size constraint of continuous valve position increment is ±2%, the weight coefficient of tracking error in the reward function is 0.7, the oil mixing loss penalty coefficient is 0.2, the energy consumption penalty coefficient is 0.1, the limit violation penalty coefficient is 0.5, the simulation environment is built based on the pipeline fluid dynamics model, and the offline training adopts the experience playback mechanism with a batch size of 128 and a learning rate of 0.0003.

[0117] Specifically, the stage operating state and reference settings are standardized and then input into the neural network of the deep reinforcement learning SAC. The neural network outputs pump frequency increment and valve position increment, and the increment values ​​are limited within the step constraint range to prevent abrupt changes. The reward function calculates the tracking error term based on the deviation between the current pump frequency, valve position and reference settings, calculates the mixing loss term based on the moving speed and density difference of the mixing interface, and calculates the energy consumption term based on pump power and valve pressure drop. When the pressure or flow exceeds the safety threshold, an over-limit penalty term is triggered. In the simulation environment, training samples are generated by randomly initializing the stage operating state and reference settings, and the action distribution is optimized by using the entropy maximization strategy to minimize multi-objective loss while satisfying the constraints of the continuous adjustment. The trained deep reinforcement learning SAC model is deployed, and the continuous adjustment is dynamically generated based on the real-time stage operating state and reference settings. It is then integrated with the rule control of the expert model library to form the final control strategy.

[0118] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0119] The deep reinforcement learning SAC algorithm takes the stage operating state and reference settings as inputs. The stage operating state includes internal and external pressure vectors, inlet and outlet flow rates, valve positions, pump frequencies, oil interface estimation, stage remaining quantity, target deviation, safety margin, and data quality labels. The reference settings include reference pump frequencies and reference valve position settings.

[0120] The action space of the SAC algorithm consists of continuous pump frequency and valve position increments. These increments are constrained by step size and range to ensure the smoothness and safety of the control action. For example, the pump frequency increment can be limited to ±0.5 Hz / s, and the valve position increment can be limited to ±1% / s.

[0121] The reward function is designed with the primary objective of minimizing the error in tracking the reference setting. Specifically, the mean squared error can be used as the main reward term.

[0122] R_tracking=-((pump frequency - reference pump frequency)) 2 +(valve position - reference valve position) 2 )

[0123] At the same time, the reward function includes terms for mixed oil loss and energy consumption penalty:

[0124] R_mixing = -k1 Estimated oil mixing loss; R_energy = -k2 Estimated energy consumption.

[0125] Where k1 and k2 are weighting coefficients.

[0126] In addition, a penalty for exceeding limits is set up, which provides a significant negative reward when parameters such as pressure and flow rate exceed the safe range:

[0127] R_limit=-1000 if limit exceeded else 0

[0128] The total reward function is:

[0129] R=R_tracking+R_mixing+R_energy+R_limit

[0130] The SAC algorithm is trained offline using a simulation environment. This simulation environment is built upon a pipeline hydraulic model and can simulate changes in parameters such as pressure and flow rate under different operating conditions. Through training with a large amount of simulation data, the SAC algorithm can learn the optimal control strategy under different states.

[0131] After training, the SAC algorithm can output continuous pump frequency increments and valve position increments as control variables based on the current operating state and reference settings. These continuous adjustment variables, combined with expert rule control, form the final control strategy.

[0132] Through the above technical solution, this application achieves continuous action space optimization control based on deep reinforcement learning. This improves control accuracy and flexibility, and enhances the ability to cope with complex operating conditions. Simultaneously, by setting penalty terms for mixing losses and energy consumption, the control strategy can simultaneously track the reference setting while considering the goals of suppressing mixing losses and energy saving. Offline training in a simulation environment avoids the risks of direct trial and error in actual operation.

[0133] In some of the solutions described above in this application, the dynamic optimization of the continuous adjustment amount may exceed the equipment safety boundary, leading to overpressure or cavitation risks. At the same time, the direct superposition of rule control and continuous adjustment amount may cause control conflicts, affecting the achievement of synchronous arrival and oil mixing suppression targets.

[0134] This application further proposes to project the safety constraints of continuous regulation quantity according to rule control into the feasible region, and then integrate it with rule control according to preset weights to form a control strategy. When there is a risk of overpressure, cavitation or interface inaccuracy in the short-term prediction based on the stage operation status, rule control is used to replace the continuous regulation quantity to ensure that the control strategy meets the requirements of synchronous arrival, oil mixing loss suppression and energy consumption constraints.

[0135] Among them, feasible domain projection ensures that the continuous adjustment amount does not exceed the physical limit allowed by the equipment by constraining the step size and range of pump frequency increment and valve position increment. The weight fusion adopts a dynamic adjustment mechanism, giving priority to the continuous adjustment amount when there is enough time left in the stage, and increasing the rule control weight when approaching the stage switch. Short-term prediction is based on the differential equation model of pressure, flow and interface velocity to predict the operating status in the next 5 to 10 seconds. The risk substitution threshold is set to 85% of the pressure upper limit or the interface position deviation exceeds 0.3%.

[0136] Specifically, after the continuous adjustment quantity is generated, it is first truncated according to the upper limit of pump frequency change rate (2Hz / s) and the upper limit of valve position change rate (5% / s). Then, a quadratic programming algorithm is used to project the adjustment quantity to the feasible solution space that satisfies the hydraulic balance equation. During the fusion process, the preset weights are dynamically adjusted according to the remaining time of the stage and the target deviation. For example, when the remaining time is greater than 300 seconds, the weight of the continuous adjustment quantity is 0.7, and the weight of the rule control is 0.3. When the remaining time is less than 60 seconds, the weight ratio is adjusted to 0.3 and 0.7. The short-term prediction module uses the fourth-order Runge-Kutta method to solve the pressure fluctuation equation. When it is predicted that the outlet pressure of a station will exceed 8.5MPa or the interface arrival time deviation will exceed ±15 seconds within the next 5 seconds, it immediately switches to the rule control mode and uses a preset pump frequency step adjustment sequence and valve holding command. Thus, while maintaining multi-objective optimization performance, the risks of equipment exceeding limits and interface loss of control are effectively avoided.

[0137] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0138] When obtaining the control strategy, the continuous adjustment quantity is first projected into the feasible region according to the safety constraints of the rule-based control. Specifically, upper and lower limit constraints are set for the pump frequency increment and valve position increment, such as pump frequency increment ±2Hz / s and valve position increment ±5% / s. Then, the projected continuous adjustment quantity and the rule-based control are fused with preset weights to form the control strategy. The fusion weights can be set to 0.7:0.3.

[0139] Furthermore, short-term predictions are made based on the stage of operation. For example, a Kalman filter is used to predict parameters such as pressure and flow rate for the next 5 to 10 minutes. When there is a risk of overpressure (e.g., predicted pressure exceeding 95% of the allowable pressure of the pipe section), cavitation risk (e.g., predicted pump inlet pressure lower than the minimum required positive suction head), or interface misalignment risk (e.g., predicted interface position deviation exceeding ±50m), rule-based control is used instead of continuous adjustment.

[0140] This ensures that the control strategy meets the requirements of synchronous arrival, suppression of oil mixing losses, and energy consumption constraints. Specifically, rule-based control ensures that the synchronous arrival deviation of each station does not exceed ±2%, oil mixing losses are controlled within 0.05%, and energy consumption does not exceed 105% of the reference value.

[0141] Through the above technical solution, this application achieves an organic combination of continuous regulation and rule-based control, improving the robustness and safety of the control strategy. By using short-term prediction and risk assessment, it promptly switches to rule-based control, effectively avoiding abnormal operating conditions such as overpressure and cavitation. Simultaneously, the integrated control strategy takes into account multiple objectives such as synchronous arrival, oil mixing losses, and energy consumption, improving the overall efficiency and stability of pipeline operation.

[0142] In some of the solutions mentioned above in this application, a cutting sequence and advance amount are generated by a mixed oil cutting strategy and the cutting valve action sequence is called. However, in the actual cutting process, due to the oil interface detection error and valve action delay, the cutting sequence deviation and advance amount may be inaccurate, resulting in interface arrival deviation or mixed oil tail exceeding expectations. At this time, the cutting parameters need to be dynamically corrected.

[0143] This application further proposes a closed-loop adjustment of the cutting timing and lead time during the mixed oil cutting process, combining oil interface detection information. When the deviation exceeds a threshold, the cutting parameters are corrected. This includes: using the stage operation status and the oil interface estimate obtained from data fusion as inputs to generate the cutting timing and lead time, with the lead time being 30-90 seconds; calling the cutting valve action sequence according to the reference setting, the action sequence including opening-holding-closing, and limiting the valve position change rate and minimum holding time to 2-5 seconds; using a short-window predictive corrector to read the oil interface detection information at a period of 1-2 seconds and predict the interface arrival time to correct the cutting timing and lead time; after the cutting is completed, the deviation is judged based on the interface arrival deviation or mixed oil trail index, and when the deviation exceeds the cutting deviation threshold, the cutting parameters are corrected, including the lead time, valve opening degree, and holding time, and the correction results are written into the mixed oil cutting strategy for the next stage.

[0144] The cutting timing and advance calculation are generated based on oil interface velocity estimation and density difference thresholds. A short-window prediction corrector reads interface detection information in real time and predicts the interface arrival time, dynamically adjusting the trigger time of the cutting action. The cutting valve action sequence adopts a three-stage operation of opening, holding, and closing. The valve position change rate is limited to no more than 5% per second, and the minimum holding time is set to 2-5 seconds to avoid frequent valve operation. The short-window prediction corrector uses oil interface displacement data within a sliding time window to predict the time for the interface to reach the target position through linear extrapolation or Kalman filtering, correcting the cutting advance. When correcting the cutting parameters, the advance parameters for the next stage are adjusted according to the interface arrival deviation and the length of the mixed oil trail. For example, when the interface arrival time delay exceeds 10 seconds, the advance is increased by 5-15 seconds; when the length of the mixed oil trail exceeds the threshold, the valve opening is increased or the holding time is extended.

[0145] Specifically, during the mixed oil cutting stage, the real-time oil interface estimate is obtained through multi-source sensor data fusion and combined with the remaining time estimate in the stage's operating state to generate the initial cutting timing and advance. The cutting valve action sequence determines the valve opening and change rate based on the flow parameters in the reference settings; for example, a lower valve position change rate is used under low flow conditions. The short-window prediction corrector analyzes the interface detection data at 1-2 second intervals to predict the actual arrival time of the interface. If the deviation between the predicted time and the initial cutting timing exceeds 5 seconds, the cutting advance is immediately adjusted. After cutting, by comparing the deviation between the actual interface arrival time and the target time, as well as the density distribution index of the mixed oil trail, it is determined whether the cutting deviation threshold has been exceeded. For example, when the interface arrival time deviation exceeds 8 seconds or the mixed oil trail length exceeds 2 meters, the cutting parameter correction logic is triggered, increasing the advance for the next stage by 10 seconds and extending the valve position holding time to 5 seconds. The corrected parameters are written into the mixed oil cutting strategy for the next stage, achieving closed-loop optimization. This process, through dynamic feedback and parameter adjustment, effectively reduces the impact of interface detection errors and valve action delays, improves cutting accuracy, and reduces oil mixing losses.

[0146] As a preferred embodiment, the specific implementation of this application is as follows: In the oil mixing and cutting stage, the oil interface detection information is obtained by multi-source sensor fusion to obtain the interface estimate. The initial cutting sequence is generated using the stage remaining amount and interface movement speed as input parameters, and the advance amount is set to 45 seconds. The cutting valve action sequence adopts a three-stage control mode, that is, the valve is opened to the preset opening degree and held for 3 seconds, and then the valve is closed at a rate of 2% per second. The short window prediction corrector collects interface position data at a 1-second cycle and predicts the time for the interface to reach the cutting point through a linear extrapolation algorithm. When the predicted time deviates from the original setting by more than 5 seconds, the advance amount parameter is automatically adjusted. After the cutting is completed, the cutting quality is determined based on the difference between the measured length of the oil mixing section and the theoretical value by the chromatograph. If the deviation exceeds the allowable threshold, a correction command is generated to increase the valve position holding time of the next stage from 3 seconds to 4 seconds and adjust the advance amount to 50 seconds.

[0147] Through the above technical solution, this application achieves dynamic closed-loop control of the mixed oil cutting process, solving the problem of cutting timing deviation caused by reliance on traditional manual experience. By combining a short-window predictive corrector with multi-source data fusion technology, the prediction accuracy of the interface arrival time is effectively improved, avoiding premature or delayed cutting caused by flow rate fluctuations. The dynamic correction mechanism of cutting parameters can adjust the valve action sequence according to actual working conditions, reducing the risk of exceeding the standard length of the mixed oil section and minimizing oil quality loss.

[0148] In some of the solutions described above in this application, there is a problem of insufficient precision in adjusting the cutting sequence and advance during the oil mixing cutting process, which leads to interface positioning deviation or oil mixing trail indicators exceeding the threshold, affecting the oil mixing loss control effect.

[0149] This application further proposes to perform closed-loop adjustment of cutting timing and advance amount by combining oil interface detection information during the oil mixing cutting process, and to correct the cutting parameters when the deviation exceeds the threshold.

[0150] The cutting timing and lead time are generated based on the stage operation status and oil interface estimation, with the lead time set to 30 to 90 seconds. The cutting valve action sequence adopts an open-hold-close mode, with limited valve position change rate and a holding time controlled between 2 and 5 seconds. The short-window predictive corrector reads oil interface detection information at a cycle of 1 to 2 seconds and predicts the interface arrival time, dynamically adjusting the cutting timing and lead time. After cutting, the deviation is determined based on the interface arrival deviation or mixed oil trail index to see if it exceeds the threshold. If it does, the cutting parameters, including the lead time, valve opening degree, and holding time, are corrected, and the correction results are written into the mixed oil cutting strategy for the next stage.

[0151] Specifically, the oil interface detection information is used to generate an interface estimate through multi-source weighted fusion. This estimate, combined with the remaining time estimate and target deviation in the stage's operational state, generates the initial cutting timing and lead time. During the execution of the cutting valve action sequence, valve position change rate and holding time limits prevent pressure fluctuations caused by frequent valve movements. The short-window prediction corrector predicts the interface arrival time using real-time interface detection data and compares it with the target time in the reference setting, dynamically adjusting the cutting lead time. For example, the lead time is shortened when the predicted interface arrives earlier, and extended when it arrives later. After cutting, if the interface arrival deviation exceeds the cutting deviation threshold, or if the mixed oil trail indicator shows an abnormal mixed oil segment length, parameter correction logic is triggered. The corrected cutting parameters compensate for time errors by adjusting the lead time, optimize valve opening and holding time to improve cutting accuracy, and are applied to subsequent stages, forming a closed-loop optimization. For example, when the detected interface arrival deviation is +5 seconds, the lead time for the next stage is reduced by 5 seconds; if the mixed oil trail length exceeds the threshold, the valve holding time is increased to 5 seconds for sufficient cutting. This process improves the accuracy and stability of mixed oil cutting through real-time data feedback and iterative parameter updates.

[0152] As a preferred embodiment, the solution of this application is implemented as follows: In the multi-stage oil transportation control process of refined oil pipelines, when a certain stage is completed, pressure, flow rate, valve position adjustment records, pump frequency adjustment sequence, and oil interface tracking data during the operation of that stage are automatically collected to form a structured operating status dataset. The calculation module generates a target deviation by comparing the difference between the stage target quantity and the actual cumulative transportation quantity. At the same time, it calculates the inter-station arrival time deviation based on the transportation synchronization error of each participating station, and generates a cutting deviation index by combining the matching degree between the oil mixing cutting action and the preset valve position sequence. Furthermore, the energy consumption monitoring module outputs the unit transportation energy consumption deviation. The correction factor generator weights and fuses the above four types of deviations according to preset weights, and outputs a correction factor with a proportional coefficient range of -5% to +8%. This coefficient is applied to the transportation allocation module of subsequent stages after passing through a limiter to dynamically adjust the planned target quantity of the unexecuted stages. After the operational data is anonymized, the pump matching rule set and the oil mixing cut-off trigger threshold are updated in the expert model library by a margin of ±0.3%. At the same time, the complete data package containing state features, control actions and execution results is input into the deep reinforcement learning training environment. The neural network parameters are updated in a sliding window manner through the experience replay buffer to complete the online incremental optimization of the strategy model.

[0153] Through the above technical solutions, this application effectively solves the problem of traditional pipeline control lacking inter-stage parameter correction and model self-updating capabilities, realizing cross-stage dynamic compensation for throughput allocation errors and reducing the cumulative effect of multi-stage execution deviations. By adaptively adjusting thresholds in the expert rule base and training the reinforcement learning model online, the control can adapt to dynamic changes in pipeline operating conditions, improving the prediction accuracy of mixed oil cutting timing and the stability of pump-valve coordinated control, forming a closed-loop optimized intelligent control system. The structured storage and directed feedback mechanism of operational data ensures the continuous evolution of the knowledge base, providing reliable data support for multi-objective optimization under complex operating conditions.

[0154] The above embodiments introduce stage objects to parse the complex multi-stage oil transportation plan into structured tasks that include target volume, allowable deviation, participating stations, oil mixing and cutting strategies, and stage switching conditions. This structured task serves as the core driver for achieving closed-loop control throughout the entire process, overcoming the limitations of existing technologies that only allow single-point adjustments. By constructing stage operating states through data fusion, and performing capacity balance optimization under equipment boundaries and hydraulic constraints, it not only achieves synchronous arrival between stations but also minimizes oil mixing losses and energy consumption constraints, enhancing global optimization capabilities. Utilizing the collaborative processing mechanism of an expert model library and deep reinforcement learning SAC, it combines the advantages of empirical rules and adaptive optimization, ensuring safety and robustness, and possessing continuous learning and dynamic optimization capabilities. In the oil mixing and cutting stage, closed-loop adjustment of timing and lead time is achieved by combining oil interface detection, reducing oil mixing losses and cutting deviations. After stage completion, deviation feedback is executed to correct capacity allocation and update the expert model library and SAC strategy, forming a continuously evolving intelligent control system. This achieves closed-loop optimization of the entire process of multi-objective operation of refined oil pipelines, combining safety, economy, and intelligence.

[0155] In another preferred embodiment based on the above embodiments, see [reference] Figure 2 As shown, this embodiment provides a multi-objective intelligent control system for refined oil pipelines based on phased operation, used to apply the above-mentioned multi-objective intelligent control method for refined oil pipelines based on phased operation, including:

[0156] The data acquisition unit is configured to create phase objects, parsing the multi-phase oil transportation plan into structured tasks. These structured tasks include phase target quantities, allowable deviations, participating stations, oil mixing strategies, and phase switching conditions.

[0157] The acquisition unit is also configured to obtain real-time operating parameters and planned parameters through data fusion. Real-time operating parameters include pressure, flow rate, valve position, pump frequency, and oil interface information, combined with the operating status during the structured task generation phase.

[0158] The first processing unit is configured to, under equipment boundary and hydraulic constraints, calculate the reference throughput of each participating station based on the stage operation status and structured tasks through throughput balance, ensuring that the station synchronously reaches the stage target throughput, minimizes mixing losses, and meets energy consumption constraints. The reference throughput generates reference settings, including reference pump frequency and reference valve position settings.

[0159] The second processing unit is configured to process the stage operation status and reference settings based on an expert model library and deep reinforcement learning SAC. Specifically, the expert model library generates rule-based controls for pump allocation, flow regulation, oil mixing, and anomaly handling, while the deep reinforcement learning SAC generates dynamically optimized continuous adjustment quantities, ultimately obtaining the control strategy.

[0160] The adjustment unit is configured to perform closed-loop adjustment of the cutting timing and advance amount during the oil-mixed cutting process, based on oil interface detection information, and to correct the cutting parameters when the deviation exceeds a threshold.

[0161] The update unit is configured to perform deviation correction of input allocation based on the stage after the stage is completed, and upload the running data to the expert model library and the deep reinforcement learning SAC for updating.

[0162] Specifically, the acquisition unit ensures the real-time performance and consistency of pressure vectors and interface estimates during the operational phase through high-precision clock synchronization and multi-source data fusion. When solving for the reference output under equipment boundary constraints, the first processing unit converts the pump characteristic curve and valve flow coefficient into a set mapping of pump frequency and valve position, providing a reference input for subsequent control. In the second processing unit, the rule-based control from the expert model library and the continuous adjustment quantities from deep reinforcement learning undergo safety constraint projection before weighted fusion. For example, when cavitation risk is predicted, the pump frequency protection strategy in the rule-based control is forcibly adopted. During the mixed oil cutting phase, the adjustment unit dynamically corrects the valve action sequence by predicting the interface arrival time. When a cutting deviation exceeding 1.5% is detected, the advance correction module is triggered to update the cutting parameters for the next phase. After the phase is completed, the update unit inputs the energy consumption deviation data into the proportional limit correction algorithm to adjust the output allocation weights for subsequent phases, while simultaneously optimizing the reward function parameters of the SAC model through the incremental training module. Each unit transmits instructions at the millisecond level through a high-speed data bus, ensuring the real-time triggering of stage switching conditions and closed-loop adjustment, ultimately forming a complete system closed loop from data acquisition, multi-objective optimization, strategy generation to dynamic updates.

[0163] As a preferred embodiment, the solution of this application is implemented as follows: The acquisition unit is configured as a phase object generation module. This module extracts the phase target quantity, participating station list, and oil mixing strategy parameters by parsing the XML format data in the multi-phase oil transportation plan, and stores them in a structured task database. The phase target quantity is received through the station PLC interface, the participating station list is obtained through the SCADA system, and the oil mixing strategy parameters are automatically generated from historical operation records. This unit further connects to the station RTU equipment via the OPC UA protocol to acquire real-time data from pressure sensors, flow meters, and oil interface detection instruments. A sliding window midpoint filtering algorithm is used to denoise the original signal, generating a phase operation status data packet containing pressure vectors, flow sequences, and interface position estimates.

[0164] The first processing unit deploys a capacity balancing optimization engine. This engine loads equipment boundary configuration files, including pump unit frequency ranges, valve opening limits, and pipeline pressure limits. After receiving the phase operation status data packets, the engine constructs a quadratic programming model with inter-station synchronization arrival time as the optimization objective. It then calls a numerical solver to calculate the reference capacity for each station and converts the capacity into a reference frequency setpoint through a pump characteristic curve conversion module, generating a set of setting instructions that includes the reference pump frequency and valve opening.

[0165] The second processing unit integrates an expert rule engine and a reinforcement learning inference module. The expert rule engine loads a predefined pump priority table and valve action sequence library. When it detects that the pipeline pressure is approaching the safety threshold, it triggers an emergency pump frequency reduction rule. The reinforcement learning inference module loads a pre-trained SAC model, inputs the current operating status and reference setpoint, outputs a pump frequency fine-tuning amount, and performs a weighted fusion with the rule engine output to generate the final control command, which is then sent to the station's actuators.

[0166] The regulating unit operates a closed-loop controller for oil mixing and cutting. This controller receives interface position data detected by an online near-infrared oil analyzer and uses a time series prediction algorithm to calculate the remaining time for the interface to reach the cutting point. When the deviation between the predicted time and the preset lead time exceeds 20 seconds, the valve action sequence is automatically adjusted and the cutting parameter configuration file is updated. At the same time, an alarm log is triggered to record the abnormal event.

[0167] The module updates the unit configuration data archiving module and model training interface. After each phase, this module extracts the control instruction sequence and result indicators from the execution process, calculates the target deviation rate and energy consumption excess, generates phase correction coefficients, and updates the configuration files for subsequent phase tasks. Simultaneously, it pushes the runtime data to the cloud training platform, triggering the expert rule base's threshold adaptive adjustment process and the incremental training task of the reinforcement learning model.

[0168] Through the above technical solutions, this application realizes the full-process automated control of multi-stage operation of refined oil pipelines. By using structured task parsing and real-time data fusion, it ensures collaborative operation of each station. By combining rule control and reinforcement learning, it optimizes and solves multi-objective conflict problems. By using a closed-loop adjustment mechanism, it improves the accuracy of oil mixing and cutting. Ultimately, it achieves the technical effects of reducing the intensity of manual intervention, reducing oil mixing losses, and ensuring the timely execution of transportation plans.

[0169] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A multi-objective intelligent control method for product oil pipeline based on stage operation, characterized in that, The method comprises the following steps: establishing a stage object, and parsing a multi-stage oil transportation plan into a structured task, the structured task comprising a stage target volume, an allowable deviation, a participating station yard, a mixed oil cutting strategy, and a stage switching condition; obtaining real-time operation parameters and planned parameters through data fusion, the real-time operation parameters comprising pressure, flow rate, valve position, pump frequency, and oil product interface information, and generating a stage operation state in combination with the structured task; under the conditions of equipment boundaries and hydraulic constraints, calculating reference throughputs of each participating station yard based on the stage operation state and the structured task, so that the station yard synchronization reaches the stage target volume, mixed oil loss is minimized, and energy consumption constraints are satisfied, and generating reference settings, the reference settings comprising reference pump frequency and reference valve position settings; processing the stage operation state and the reference settings based on an expert model library and a deep reinforcement learning SAC, wherein rules for pump matching, flow rate adjustment, mixed oil cutting, and abnormality disposal are generated from the expert model library, dynamic optimization continuous adjustment amounts are generated from the deep reinforcement learning SAC, and a control strategy is obtained according to the rules and the continuous adjustment amounts; in the process of mixed oil cutting, the cutting timing and advance amount are closed-loop adjusted in combination with oil product interface detection information, and cutting parameters are corrected when the deviation exceeds a threshold value; after completion of a stage, the throughputs of subsequent stages are adjusted according to stage execution deviations, and stage operation data is uploaded to the expert model library and the deep reinforcement learning SAC for updating; under the conditions of equipment boundaries and hydraulic constraints, when the reference throughputs of each participating station yard are calculated based on the stage operation state and the structured task, and the reference settings are generated, the following steps are included: taking the stage operation state and the structured task as inputs, taking the minimum station yard residual arrival time deviation as a synchronization index, and taking mixed oil loss and energy consumption as parallel targets; under the conditions of equipment boundaries and hydraulic constraints, the reference throughputs of each participating station yard are solved by using quadratic programming, the equipment boundaries comprising pump frequency ranges and variation rates, valve position ranges and variation rates, and the hydraulic constraints comprising upper limits of pipe segment pressure, minimum required net positive suction head, and upper limits of pipe flow rate; the reference throughputs are mapped to reference pump frequency and reference valve position settings to form the reference settings according to pump characteristic curves and valve flow rate coefficients; the expert model library stores pump matching models, flow rate adjustment models, mixed oil cutting models, and abnormality disposal models in the form of decision trees or rule sets, takes the stage operation state and the reference settings as inputs, outputs rule controls according to pre-set triggering output rules, the rule controls comprising pump frequency setting increments, valve position setting increments, and action sequences, and performs boundary checking on stage switching conditions and allowable deviations; when there is a rule conflict, the rule is decided in the order of safety priority > mixed oil cutting > flow rate adjustment > pump matching.

2. The multi-objective intelligent control method for product oil pipeline based on stage operation according to claim 1, characterized in that, when the stage object is established, and the multi-stage oil transportation plan is parsed into a structured task, the following steps are included: performing consistency checking on the stage target volume and the participating station yard according to equipment boundaries and hydraulic constraints, and when it is not feasible, decomposing or merging stages and redistributing stage target volumes according to priorities; generating an allowable deviation based on quantile statistics and risk levels of historical stage tracking errors, the allowable deviation being 0.5% to 1.5%. The cutting timing and advance of the mixed oil cutting strategy are determined based on the interface speed estimation and the density difference threshold, and a cutting valve action sequence is given; The stage switching condition is jointly triggered by the stage cumulative amount reaching the stage target amount minus the allowable deviation, the interface reaching the target position, and meeting the safety state, and a hysteresis window is set.

3. The multi-objective intelligent control method for product oil pipeline based on stage operation according to claim 2, characterized in that, When the real-time running parameters and the planned parameters are obtained through data fusion, and the stage running state is generated in combination with the structured task, it includes: The real-time running parameters and the planned parameters are collected by the RTU through OPC UA, and the clock is synchronized using PTP, so that the clock deviation is not greater than 100 milliseconds; The abnormal values of the pressure, flow, valve position, pump frequency, and oil product interface information are removed, and 5-11 point sliding median denoising is performed to complete the unit and dimension unification; Spline interpolation is performed on the missing interval of not more than 60 seconds; the sensor is calibrated for zero and span; the oil product interface information is multi-source weighted fusion and the interface estimation value is output; The stage target amount, allowable deviation, participating station, and stage switching condition are mapped to a unified time axis, and the stage remaining amount, target deviation, and remaining time estimation of inter-station synchronous arrival are calculated. The stage running state is generated using Kalman filtering with a period of 1-5 seconds, and the stage running state includes the pressure vector inside and outside the station, the in-out station flow, the valve position, the pump frequency, the oil product interface estimation, the stage remaining amount, the target deviation, the safety margin, and the data quality mark.

4. The multi-objective intelligent control method for product oil pipeline based on stage operation according to claim 1, characterized in that, The deep reinforcement learning SAC takes the stage running state and the reference setting as input; the action is a continuous pump frequency increment and a valve position increment, which is constrained by the step size and the range; the reward is mainly to minimize the error of tracking the reference setting, while adding mixed oil loss and energy consumption penalties and out-of-limit penalties; the simulation environment is used for offline training and the continuous adjustment amount is output.

5. The multi-objective intelligent control method for product oil pipeline based on stage operation according to claim 4, characterized in that, When the control strategy is obtained, it includes: First, the continuous adjustment amount is projected to the feasible region according to the safety constraints of the rule control, and then it is fused with the rule control according to the preset weight to form the control strategy; When there is a risk of overpressure, cavitation, or interface misalignment in the short-term prediction based on the stage running state, the rule control is replaced by the continuous adjustment amount to ensure that the control strategy meets the requirements of synchronous arrival, mixed oil loss suppression, and energy consumption constraints.

6. The multi-objective intelligent control method for product oil pipeline based on stage operation according to claim 5, characterized in that, In combination with the oil product interface detection information, the cutting timing and advance are adjusted in a closed loop, and when the deviation exceeds the threshold, the cutting parameters are corrected, including: The stage running state and the oil product interface estimation obtained by data fusion are input to generate the cutting timing and advance, and the advance is 30-90 seconds; The cutting valve action sequence is called according to the reference setting, and the action sequence includes opening-holding-closing, with a valve position change rate and a minimum holding time of 2-5 seconds; The short window predictor reads the oil product interface detection information and predicts the interface arrival time at a period of 1-2 seconds, and corrects the cutting timing and advance; After the cutting is completed, the deviation is determined based on the interface arrival deviation or the mixed oil tail index, and when the deviation exceeds the cutting deviation threshold, the cutting parameters including the advance, the valve opening degree, and the holding time are corrected, and the correction result is written into the mixed oil cutting strategy of the next stage.

7. The multi-objective intelligent control method for product oil pipeline based on stage operation according to claim 6, characterized in that, After the stage is completed, the stage execution deviation is corrected according to the stage, the running data is uploaded to the expert model library and the deep reinforcement learning SAC for updating, and the running data is uploaded to the expert model library and the deep reinforcement learning SAC for updating, including: The stage running state, reference setting, control strategy and stage cumulative quantity are summarized, the stage execution deviation is calculated, and the stage execution deviation includes target deviation, synchronization arrival deviation, mixed oil cutting deviation and energy consumption deviation; Based on the stage execution deviation, a correction factor is generated, the stage target quantity distribution and the reference setting of the subsequent stage are proportionally limited and corrected, and the consistency of the stage switching condition and the allowed deviation is maintained; The running data including the stage running state, the reference setting, the rule control, the continuous adjustment quantity, the control strategy and the stage execution deviation are uploaded to the expert model library and the deep reinforcement learning SAC; wherein the expert model library limits and updates the rule threshold and the action timing, and the deep reinforcement learning SAC takes the running data as an experience replay sample for incremental training.

8. A multi-objective intelligent control system for stage-based operation of a product oil pipeline, for applying the multi-objective intelligent control method for stage-based operation of a product oil pipeline according to any one of claims 1 to 7, characterized in that, Including: The acquisition unit is configured to establish a stage object, parse a multi-stage oil transportation plan into a structured task, and the structured task includes a stage target quantity, an allowed deviation, a participating station, a mixed oil cutting strategy and a stage switching condition; The acquisition unit is also configured to obtain real-time running parameters and plan parameters through data fusion, the real-time running parameters include pressure, flow, valve position, pump frequency and oil product interface information, and generate a stage running state in combination with the structured task; The first processing unit is configured to calculate the reference flow of each participating station based on the stage running state and the structured task under the conditions of device boundary and hydraulic constraint, so that the inter-station synchronization arrival stage target quantity, the mixed oil loss minimization and the energy consumption constraint are satisfied, and the reference setting is generated, the reference setting includes the reference pump frequency and the reference valve position setting; The second processing unit is configured to process the stage running state and the reference setting based on the expert model library and the deep reinforcement learning SAC, wherein the rule control of pump distribution, flow adjustment, mixed oil cutting and abnormal handling is generated by the expert model library, and the dynamic optimization continuous adjustment quantity is generated by the deep reinforcement learning SAC, and finally the control strategy is obtained; The adjustment unit is configured to adjust the cutting timing and the advance quantity in combination with the oil product interface detection information during the mixed oil cutting process, and correct the cutting parameters when the deviation exceeds the threshold; The update unit is configured to correct the flow distribution according to the stage execution deviation after the stage is completed, and upload the running data to the expert model library and the deep reinforcement learning SAC for updating.

Images