Oil pumping unit well time-sharing optimization design method and system based on edge calculation

Abstract

The invention discloses a rod pumped well time-sharing optimization design method and system based on edge calculation, and relates to the field of rod pumped well time-sharing optimization, and the method comprises the following steps: S1, constructing an edge sensing layer, and collecting data in real time; s2, edge layer data extraction and working condition diagnosis; s3, time-of-use electricity price response and working condition prediction; s4, generating an edge end real-time control strategy; s5, edge-cloud collaborative reinforcement learning optimization is carried out; and S6, performing dynamic evaluation and strategy correction. According to the time-sharing optimization design method and system for the rod-pumped well based on edge calculation, self-adaptive control over high-stroke-frequency yield increase in the valley period and low-stroke-frequency peak avoidance in the peak period can be achieved, and the phenomenon of empty pumping in the high electricity price period can be avoided; when the liquid supply is sufficient in the valley period, the stroke frequency can be increased, the liquid production capacity in the low-electricity-price period can be maximized, and the problem that timing control is disjointed with the actual working condition is fundamentally solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of well time optimization of pumping units, in particular to a pumping unit well time optimization design method and system based on edge computing. BACKGROUND

[0002] In the field of oilfield digitization, the energy consumption optimization of pumping unit wells has long relied on artificial experience and static control strategies. Most existing technologies use cloud-based centralized computing to analyze well site data, but are limited by network transmission delays and cannot meet real-time response needs. Local PLC control schemes, while reducing delays, lack online learning capabilities and are difficult to handle complex tasks such as time series data prediction and multi-objective optimization. Especially in the context of time-of-use electricity pricing, traditional methods do not establish a joint decision-making model for electricity pricing periods, pump efficiency, and fault states, resulting in insufficient production during off-peak hours and excessive energy consumption during peak hours. In addition, fault diagnosis still relies on manual analysis of dynamometer cards, and the real-time inference capabilities of edge computing nodes are not utilized, resulting in high false positive rates. How to achieve closed-loop optimization of data acquisition, working condition diagnosis, and dynamic strategy generation through a distributed computing architecture is a core challenge for current digital data processing technologies in industrial control.

[0003] Therefore, it is necessary to propose a pumping unit well time optimization design method and system based on edge computing to solve the above problems. SUMMARY

[0004] The main purpose of the present application is to provide a pumping unit well time optimization design method and system based on edge computing, which can effectively solve the problems in the background art.

[0005] To achieve the above purpose, the technical scheme adopted by the present application is as follows: A pumping unit well time optimization design method based on edge computing, comprising the following operation steps: S1: Edge perception layer construction and real-time data acquisition, used to construct a high-precision, low-delay data acquisition network covering key operating parameters of pumping unit wells; S2: Edge layer data extraction and working condition diagnosis, based on the data calculated by S1, the pump efficiency index of the pumping unit is calculated and the fault is diagnosed, providing a basis for strategy formulation; S3: Time-of-use electricity pricing response and working condition prediction, combining the electricity pricing period and the fullness prediction of the pumping unit pump to generate a production strategy prototype; S4: Real-time control strategy generation at the edge, based on the pump fullness of the pumping unit and the electricity pricing period, dynamically adjusting the pumping unit stroke to achieve increased production during off-peak hours and reduced energy consumption during peak hours; S5: Edge-cloud collaborative reinforcement learning optimization, continuously optimizing the control strategy through multi-objective reinforcement learning; S6: Dynamic evaluation and strategy modification, used for continuously tracking the optimization effect of the strategy and self-adaptively adjusting.

[0006] Preferably, in S1, the data acquisition is performed by a three-phase current sensor, a load displacement sensor, and a pressure sensor. The three-phase current sensor is installed at the three-phase input terminal of the motor of the pumping unit to collect the effective value, peak value, and phase difference of the current at a frequency of 5-10 times per stroke, and to synchronously calculate the motor power of the pumping unit. The load displacement sensor is installed axially on the load-bearing rod of the hanger at the wellhead of the pumping unit to monitor the load at the suspension point and the displacement of the polished rod in real time. The pressure sensor is embedded through the threaded interface at the bottom of the tubing of the pumping unit to monitor the change in the dynamic liquid level and to calculate the dynamic liquid level change rate based thereon. The above data is pre-processed by the edge computing node of the edge RTU, and is transmitted to the edge node. The edge RTU is also provided with a CNN model.

[0007] Preferably, in S2, the following steps are specifically included: S201: Calculation of the pump fullness of the pumping unit, the high-frequency noise of the load displacement sensor is removed by using a sliding average filter, the load displacement data and the motor current data of the pumping unit are synchronized to the same stroke cycle by using a timestamp alignment technology, and the theoretical stroke area is calculated based on a formula: ; wherein is the theoretical stroke area; is the cross-sectional area of the plunger; is the set stroke length of the pumping unit; a contour detection algorithm based on Canny edge detection is used to identify the closed curve of the indicator diagram, the effective area of the actual liquid discharge stroke is calculated by trapezoidal integration of the area inside the closed curve, and the actual liquid discharge area is extracted therefrom , and the fullness is quantified based thereon: ; S202: Peak value detection is performed on the data of the three-phase current sensor, the maximum values of the A, B, and C three-phase currents collected by the three-phase current sensor are taken as the upstroke and downstroke peak values, and are denoted as and , respectively. The sampling period is synchronized with the stroke, and a group of peak value data is recorded for each stroke. The current balance degree is calculated based thereon: ; When , it is in a balanced state. S203: Lightweight fault diagnosis implementation, based on edge RTU deployment CNN model to identify and warn the pumping unit fault type, CNN model is used to output the fault type of pumping unit and its confidence, the fault type includes normal, insufficient supply, valve leakage, gas lock, when the confidence of a single fault type is > 90%, and more than one fault type is output at the same time, the alarm is triggered.

[0008] Preferably, in S201, when the pumping unit pump is in normal supply, when the pumping unit pump needs to be warned for liquid supply, and when the pumping unit pump is judged to be insufficient supply.

[0009] Preferably, in S203, the network architecture of the CNN model includes: Input layer: 256x256 pixel gray dynamometer diagram; output layer: output fault probability type and probability distribution of normal, insufficient supply, valve leakage and gas lock; Training data set 100,000 labeled dynamometer diagrams, including normal working condition: 40,000, i.e. , ; insufficient supply: 30,000, i.e. ; valve leakage: 20,000, i.e. sawtooth fluctuation appears in the dynamometer diagram; gas lock: 10,000, i.e. the dynamometer diagram presents pear-shaped distortion; Second, edge RTU collects 1 set of dynamometer diagram every time, which is automatically scaled to 256x256 pixels, and performs convolution operation through FPGA acceleration unit, and then outputs the fault type and confidence of pumping unit; Every 1000 new dynamometer diagrams are accumulated and diagnosed, the model fine-tuning is automatically started, and when the confidence of continuous 5 times diagnosis is < 70%, the incremental learning of CNN model is forced to trigger.

[0010] Preferably, in S3, it specifically includes the following steps: S301: Electricity price period division, the electricity price is divided into three periods of 08:00-12:00, 1.2 yuan / kWh peak, 12:00-18:00, 0.7 yuan / kWh flat, 00:00-06:00, 0.3 yuan / kWh valley, edge RTU synchronizes the power grid price table in real time through 4G, and the local cache is valid for 7 days; S302: Pump fullness prediction model implementation, according to the improved gray prediction model: ; The prediction process is: for accumulation to get ; solve , by least square method; introduce , compensating for dynamic liquid level mutation error, predicting future 2 hours trend; when the dynamic liquid level change rate > 10 m / h, automatically adjusting the power index: ; wherein is the historical pump fullness sequence; , is the gray model coefficient, wherein is the development coefficient, is the gray action amount; is the power correction coefficient; is the power index; is the cumulative generation sequence, obtained by accumulating ; is the time step; is the natural constant.

[0011] Preferably, in the S4, the following steps are specifically included: S401: The policy triggering condition includes: Increase production during valley period: When in valley electricity price period and pump fullness > 80%, the edge RTU starts the stroke lifting mechanism; Reduce consumption during peak period: Peak electricity price period and < 50%, the pumping unit motor is switched to one of the low-speed mode of 1 / min and shutdown; S402: The implementation process specifically includes: Stroke adjustment range, the rated stroke reference value is 6 / min, and the maximum lifting during the valley period is 130% of the rated value; the low-speed mode is set to 1 / min, and the pumping unit motor inertial slides to stop when shutdown; PID control: dynamically adjust the frequency converter frequency of the pumping unit through the PID controller; Real-time feedback: collect pumping unit motor current and speed data every stroke, if the stroke fluctuation exceeds the set value ± 5%, the edge RTU automatically fine-tunes the frequency converter frequency, and the adjustment step is 0.5 Hz; S403: Current balance degree closed-loop control strategy, prepare the beam balance block, install it at the tail of the pumping unit beam, and realize horizontal position adjustment through the stepping motor driven screw transmission mechanism, specifically: Real-time monitoring of current balance degree , automatically adjusting the position of the beam balance block, and maintaining the current balance degree within 80%-110%; when: Less than 80%, it is judged that the downstroke current is too large and the beam balance block counterweight is insufficient; If greater than 110%, it is determined that the upstroke current is too large, and the balance block is too heavy, and the edge RTU starts the adjustment mechanism: The edge RTU outputs a signal to drive the stepper motor, which drives the walking beam balance block to move horizontally along the tail of the walking beam through a lead screw; The adjustment logic is as follows: When greater than 110%, the walking beam balance block moves away from the wellhead, increasing the upstroke weight and reducing the upstroke current; when Less than 80, the walking beam balance block moves towards the wellhead, reducing the downstroke load and increasing the downstroke current; After each adjustment, the current peak of three strokes is continuously monitored, and if It does not return to the target range, secondary adjustment is started after 1 minute, and after a maximum of 3 attempts, manual intervention alarm is triggered; S404: intermittent cycle adaptive strategy, calculate the fullness Every ten minutes, the decline rate is greater than 0.5% / minute: it is determined that the liquid supply is sufficient, and the oil pumping is extended to =40%; when the decline rate is less than 0.2% / minute: it is determined that the liquid supply is insufficient, and the machine is stopped in advance; S405: dynamic correction, wake up the pressure sensor every 30 minutes during shutdown, monitor the dynamic fluid level recovery, if the recovery rate is greater than 0.5% / minute, the shutdown can be ended in advance, if the dynamic fluid level does not change continuously for 2 times, the shutdown time is automatically extended by 20%, to avoid frequent start and stop.

[0012] Preferably, S5, specifically, the edge RTU collects the state data of the pumping unit and performs control actions, and uploads the state transition record to the cloud, the cloud optimizes the strategy model through the deep deterministic policy gradient algorithm, and then issues it to the edge RTU for deployment, forming a closed-loop iteration of data collection, strategy optimization, and control execution, wherein the state vector of the model includes pump fullness, current balance, electricity price, and electricity price period; the action vector includes stroke, balance, and start-stop, wherein the stroke is adjusted in the range of -50% to +30%; the balance of the walking beam balance block is adjusted in the range of -10 cm to +10 cm; start-stop takes the value of 0 and 1, 0 means the pumping unit motor is stopped, and 1 means the pumping unit motor is running; The weighted reward term guides the policy to optimize in the direction of high yield, low consumption, balance, and valley electricity use, specifically: Yield reward: the higher the ratio of actual liquid production to target liquid production, the higher the reward, which is used to encourage stable production; Energy consumption penalty: the higher the ratio of actual energy consumption to rated energy consumption, the greater the penalty, which is used to suppress high energy consumption; Balance penalty: the greater the deviation of balance from 100%, the greater the penalty, which is used to promote motor load balance; Off-peak reward: the higher the off-peak production time ratio, the higher the reward, which is used to increase production by using low-cost electricity; Dynamic weight adjustment: the cloud updates the weight according to the oilfield production target every day.

[0013] Preferably, in S6, the following steps are specifically included: S601: Evaluation index, specifically power saving rate: compare the total energy consumption of the pumping unit before and after optimization, automatically calculate daily and synchronize to the cloud database; Liquid production fluctuation monitoring: deviation of daily average liquid production for 3 consecutive days from historical average; Current balance Compliance rate, daily statistics of current balance Time length within 80%-110% range, target value > 90%; S602: Strategy correction, when the liquid production fluctuation monitored for 3 consecutive days is > 5%, collect the pumping unit state data for the past 7 days and mark the abnormal period, including sudden drop of dynamic liquid level, fault alarm period; based on abnormal data, recalculate the state transition probability, identify the strategy failure scenario; use Bayesian optimization algorithm to adjust the learning rate and discount factor parameters of the deep deterministic policy gradient algorithm model, to improve the adaptability of the model to abnormal working conditions; If the power saving rate has not improved for 15 consecutive days and is < 25%, trigger cloud model retraining, focus on optimizing time-of-use electricity price response strategy; When the liquid production fluctuation is > 5% for 3 consecutive days When the balance compliance rate is < 80%, analyze The working condition characteristics of the period that does not meet the standard, including peak valley period distribution, pump fullness interval, when it is found that the valley period has high stroke Frequent over-limiting, then it is determined that the coordination strategy of stroke adjustment and beam balance block adjustment is invalid; when retraining the deep deterministic policy gradient algorithm in the cloud, increase the weight of the 5% and Related reward items; optimize the coordination logic of stroke adjustment and beam balance block adjustment, when the stroke is improved by more than 20%, automatically trigger the beam balance block pre-adjustment, which is used to compensate for load changes in advance; increase the priority of beam balance block adjustment when When the compliance rate is < 70% and the liquid production fluctuation is < 5%, preferentially execute Optimized strategy correction.

[0014] A time-sharing optimization design system for pumping wells based on edge computing includes an edge sensing and data acquisition module, an operating condition diagnosis and feature extraction module, a time-sharing electricity price response and prediction module, an edge-end real-time control execution module, an edge-cloud collaborative reinforcement learning module, a security and compliance assurance module, and a dynamic evaluation and strategy correction module. The edge sensing and data acquisition module is used to collect real-time operating parameters of the pumping well and provide real-time operating data, with a sampling frequency of 5-10 seconds per time and an accuracy of ±0.5%FS. The working condition diagnosis and feature extraction module outputs the pump fullness, current balance and fault type of the pumping unit based on the data collected by the edge sensing and data acquisition module. This is used to quantitatively evaluate the operating status of the oil well and identify pumping unit faults such as insufficient fluid supply and valve leakage, providing a basis for production strategy. The time-of-use electricity price response and prediction module divides the time period into peak, flat and valley based on the electricity price. At the same time, the edge RTU synchronizes the grid electricity price in real time and predicts the next 2 hours based on the improved grey prediction model. trends; The edge-side real-time control execution module generates minute-level control instructions based on strategies to adjust the operating parameters of the pumping unit. These strategies include stroke optimization: adjusting the motor speed of the pumping unit through a frequency converter, increasing the stroke rate to 130% of the rated value during off-peak hours; current balance closed-loop: adjusting the position of the walking beam balance block through a stepper motor; and intermittent pumping cycle adaptation: dynamically adjusting the pumping duration and downtime based on the fluid supply rate. The edge-cloud collaborative reinforcement learning module realizes the autonomous evolution of strategies through data-driven, reducing the overall energy consumption by 25-42%. Its core framework includes 、 , electricity price, and electricity price period; the action vector is composed of the stroke frequency adjustment, the adjustment of the walking beam balance block, and the start and stop of the pumping unit motor; Cloud-based deep deterministic policy gradient algorithm optimization strategy, with reward function coupled with output, energy consumption, balance, and valley power ratio; Transfer learning framework, reducing cold start time of new wells by 70%; The security and compliance assurance module is used to ensure system operation security and data transmission compliance, including edge video analysis, data encryption and zero-trust network. Edge video analysis is based on a wide-angle infrared camera and YOLOv5s model to detect violations; data encryption is based on the national secret SM4 algorithm and HMAC-SHA256 integrity verification; the zero-trust network uses two-way certificate authentication at the edge node to block unauthorized access; The dynamic evaluation and strategy correction module corrects the strategy based on the power saving rate, liquid production fluctuation monitoring, and balance compliance rate statistics as evaluation indicators to ensure that the pumping unit maintains a high power saving rate and production stability for a long time.

[0015] Compared with the prior art, the present application provides an edge computing-based pumping unit well time-sharing optimization design method and system, which has the following beneficial effects: The edge computing-based pumping unit well time-sharing optimization design method and system can realize adaptive control of high-stroke yield increase in valley period and peak avoidance in peak period through real-time collection of pump fullness, dynamic liquid level depth and other parameters of the pumping unit by the edge perception layer, dynamic prediction of future 2-hour working conditions of the pumping unit by combining the improved gray prediction model, automatic triggering of low-speed mode or shutdown when the pump fullness is detected to be lower than 40%, and avoidance of air pumping in high electricity price period based thereon. When the liquid supply is sufficient in the valley period, the stroke is increased to 130% of the rated value, so that the liquid production in the low electricity price period can be maximized, and the problem of disconnection between time control and actual working conditions can be fundamentally solved.

[0016] The edge computing-based pumping unit well time-sharing optimization design method and system can realize pre-processing, working condition diagnosis and control instruction generation of data on the edge end through the use of FPGA hardware acceleration edge RTU, effectively shorten the calculation delay, and realize real-time monitoring of the current peak value ratio by the edge RTU in the current balance degree closed-loop control. When the balance degree exceeds the range of 80% to 110%, the position of the walking beam balance block is immediately adjusted by driving the stepper motor, so that the invalid energy consumption ratio can be effectively reduced.

[0017] The edge computing-based pumping unit well time-sharing optimization design method and system can integrate yield, energy consumption, balance degree and valley electricity production ratio into a unified reward function, realize multi-objective collaborative optimization, and realize real-time execution of stroke adjustment and balance block optimization on the edge end. The cloud side continuously optimizes the strategy parameters based on historical data to form a real-time control-strategy evolution closed loop.

[0018] The edge computing-based pumping unit well time-sharing optimization design method and system first dynamically binds the gray prediction model and the time-of-use electricity price, realizes precise peak avoidance and efficient use of valleys through the coupling decision mechanism of electricity price and fullness, forcibly stops the machine to store liquid in the peak period 2 hours before the predicted valley period starts to ensure efficient production in the valley period, and automatically increases the stroke to the upper limit if the liquid supply is sufficient in the valley period. This mechanism breaks the rigid mode of fixed-time start and stop, and can effectively improve the daily average power saving rate of a single well. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is a flowchart of the present application. DETAILED DESCRIPTION

[0020] In order to make the technical means, creative features, purposes and effects of the present application easy to understand, the present application will be further described below in combination with specific embodiments.

[0021] Embodiment one: As Figure 1 shown, an edge computing-based pumping unit well time-sharing optimization design method includes the following operation steps: S1: edge perception layer construction and real-time data collection, used to construct a high-precision, low-delay data collection network covering key operating parameters of pumping unit wells, through three-phase current sensors, load displacement sensors, and pressure sensors to collect data, wherein the three-phase current sensor is installed at the three-phase input end of the motor of the pumping unit, and the effective value, peak value, and phase difference of the current are collected at a frequency of 5-10 seconds per stroke, and the motor power of the pumping unit is calculated synchronously; the load displacement sensor is installed on the axial of the load-bearing rod of the suspension device at the wellhead of the pumping unit, used to monitor the load and displacement of the suspension point in real time; the pressure sensor is embedded through the threaded interface at the bottom of the oil pipe of the pumping unit, used to monitor the change of the dynamic liquid level and calculate the dynamic liquid level change rate based on the same; The edge RTU also deploys a CNN model for data preprocessing of the edge computing nodes of the edge RTU and transmitting the data to the edge nodes.

[0022] S2: edge layer data extraction and working condition diagnosis, based on the data collected in real time in S1 to calculate the pump efficiency index of the pumping unit and diagnose faults to provide a basis for strategy formulation, including the following steps: S201: calculation of pumping unit pump fullness, using sliding average filtering to remove high-frequency noise of the load displacement sensor, synchronizing the load displacement data and pumping unit motor current data to the same stroke cycle through timestamp alignment technology, and calculating the theoretical stroke area based on the formula: ; Wherein is the theoretical stroke area; is the cross-sectional area of the plunger; is the set stroke length of the pumping unit; The contour detection algorithm based on Canny edge detection is used to identify the closed curve of the indicator diagram, and the effective area of the actual liquid discharge stroke is calculated by trapezoidal integration of the inner region of the closed curve to extract the actual liquid discharge area, and the fullness is quantified based on this: ; When is normal for the pumping unit pump; when , the pumping unit pump needs to be warned for liquid supply; when , the pumping unit pump is judged to be insufficient for liquid supply; S202: peak value detection on three-phase current sensor data, taking the maximum values of A, B, and C three-phase currents collected by the three-phase current sensor as the upstroke and downstroke peak values, respectively, denoted as and , the sampling period is synchronized with the stroke, and a set of peak data is recorded for each stroke, and the current balance degree is calculated based on this: ; When , it is in a balanced state, and when it exceeds the range, the adjustment instruction of the beam balance block is triggered; S203: Lightweight fault diagnosis implementation, based on the CNN model deployed on the edge RTU to identify and warn the fault type of the pumping unit, the CNN model is used to output the fault type of the pumping unit and its confidence, the fault type includes normal, insufficient liquid supply, valve leakage, and gas lock, when the confidence of a single fault type is > 90%, and more than one fault type is output at the same time, an alarm is triggered, the network architecture of the CNN model includes: Input layer: 256x256 pixel gray dynamometer diagram; output layer: output fault probability type and probability distribution of normal, insufficient liquid supply, valve leakage, and gas lock; The training data set contains 100,000 labeled dynamometer diagrams, including normal working conditions: 40,000, i.e. , ; insufficient liquid supply: 30,000, i.e. ; valve leakage: 20,000, i.e. the dynamometer diagram shows sawtooth fluctuation; gas lock: 10,000, i.e. the dynamometer diagram shows pear-shaped distortion; Second, the edge RTU automatically scales to 256x256 pixels for each set of dynamometer diagrams collected, and performs convolution operation through the FPGA acceleration unit, and then outputs the fault type and confidence of the pumping unit. When the CNN model infers the fault type, enable dynamic algorithm power distribution: turn off FPGA acceleration in normal working conditions, and enable full power mode for fault diagnosis; Every 1000 new dynamometer diagrams are diagnosed, the model fine-tuning is automatically started, and when the confidence is < 70% for 5 consecutive diagnoses, the incremental learning of the CNN model is triggered, wherein the fine-tuning process is: randomly extract 200 samples containing fault samples from new data, and label them through active learning algorithm; use transfer learning, fix the first 3 layers of the pre-trained model, and only fine-tune the latter classifier; training rounds: 10 rounds, learning rate 10 -5 , to prevent overfitting.

[0023] S3: Time-of-use price response and working condition prediction, combining the time-of-use price period and the fullness prediction of the pumping unit to generate a production strategy prototype, including the following steps: S301: Time-of-use price division, divide the price into three periods: 08:00-12:00, 1.2 yuan / kWh peak, 12:00-18:00, 0.7 yuan / kWh flat, 00:00-06:00, 0.3 yuan / kWh valley, the edge RTU synchronizes the power grid price table in real time through 4G, and the local cache is valid for 7 days; S302: Pump fullness prediction model implementation, according to the improved grey prediction model: ; Prediction process: to Cumulative generation ; solve , ; introduce , compensate for the sudden change of liquid level error, predict the trend of the next 2 hours ; When the rate of change of liquid level is > 10m / h, automatically adjust the power index: ; Among them is the historical pump fullness sequence, sampled every 10 minutes, and 24 hours of data are retained; , is the grey model coefficient, in which is the development coefficient, which is used to reflect the rate trend of the change of pump fullness, is the grey action, which is used to characterize the comprehensive action of external factors affecting the change of pump fullness, such as the comprehensive action of the natural rise of the liquid level; is the power correction coefficient, in which is used to dynamically compensate for the sudden change of liquid level error, and the historical error is dynamically generated, which is specifically based on the average of the prediction error of the last three time steps; is the power index, which is used to control the rate of change over time, the typical value is 0.5~1.5, the default value is 1.0; is the cumulative generation sequence, which is obtained by accumulating ; is the time step, which is trained based on 24 hours of historical data, and is used to predict the next 12 time steps; is a natural constant, which is approximately equal to 2.71828 here, and is used to represent the exponential decay or growth trend of pump fullness over time.

[0024] S4: Edge real-time control strategy generation, based on the pump fullness of the pumping unit and the dynamic adjustment of the pumping unit stroke of the pumping unit according to the electricity price period, to realize the increase of production in valley period and the reduction of consumption in peak period, which specifically includes the following steps: S401: Strategy triggering conditions include: Valley period production increase: when in valley electricity price period and pump fullness > 80%, the edge RTU starts the stroke lifting mechanism; Peak period consumption reduction: peak electricity price period and < 50%, the pumping unit motor is switched to one of the low-speed mode of 1 time / min and shutdown. S402: The implementation process specifically includes: The stroke adjustment range is 6 strokes per minute, and the maximum increase in the trough period is 130% of the rated value. The low-speed mode is set to 1 stroke per minute, and the pumping unit motor inertially slides to a stop during shutdown, with a sliding time of about 15 seconds, which is used to reduce mechanical impact; PID control: dynamically adjust the frequency of the pumping unit's frequency converter through the PID controller; Real-time feedback: collect pumping unit motor current and speed data every stroke, and if the stroke fluctuation exceeds the set value ±5%, the edge RTU automatically fine-tunes the frequency converter frequency, with a step size of 0.5 Hz; S403: Current balance degree closed-loop control strategy, prepare the beam balance block, install it at the tail of the pumping unit beam, and adjust the horizontal position of the beam balance block through the stepper motor driven screw transmission mechanism. By adjusting the position of the beam balance block, the weight difference of the pumping unit's sucker rod string and plunger in the up and down strokes is offset, the current peak of the pumping unit motor in the up and down strokes is made consistent, and the energy consumption is reduced. Specifically: Real-time monitoring of current balance degree Automatic adjustment of beam balance block position to maintain current balance degree Within 80%-110%, when: Less than 80%, it is judged that the downstroke current is too large, and the beam balance block is not heavy enough; Greater than 110%, it is judged that the upstroke current is too large, and the balance block is too heavy, and the edge RTU starts the adjustment mechanism: The edge RTU output signal drives the stepper motor, which drives the beam balance block to move horizontally along the tail of the beam through the screw transmission, with an adjustment accuracy of ±2 cm, and a single maximum adjustment amount not exceeding 10% of the full stroke of the beam balance block, to avoid system oscillation caused by over-adjustment; The adjustment logic is: when > 110%, the beam balance block moves away from the wellhead to increase the upstroke weight and reduce the upstroke current; when < 80, the beam balance block moves towards the wellhead to reduce the downstroke load and improve the downstroke current; After each adjustment, the current peak of 3 strokes is continuously monitored, and if It does not return to the target range, secondary adjustment is started after 1 minute, and after a maximum of 3 attempts, manual intervention alarm is triggered; S404: Adaptive strategy for intermittent pumping period, calculate the fullness The decline rate every ten minutes, when the decline rate > 0.5% / min: it is judged that the liquid supply is sufficient, and the pumping is extended to = 40%; when the descending rate < 0.2% / min: it is judged that the liquid supply is insufficient, and the machine is stopped in advance, and the stopping time is based on the average recovery time of the well history, which refers to the average value of the time required for the dynamic liquid level to rise to make the fullness prediction value reach the preset threshold from the stopping time in the past 30 days of the well; the higher the initial fullness, the longer the stopping time, and when the initial fullness is 70%, the stopping time is 1.6 times of the average recovery time of the well history; S405: dynamic correction, wake up the pressure sensor every 30 minutes during shutdown to monitor the rising of the dynamic liquid level, if the rising rate > 0.5% / min, the shutdown can be ended in advance, if there is no change in the dynamic liquid level for two consecutive times of monitoring, the shutdown time is automatically extended by 20% to avoid frequent start-stop.

[0025] S5: edge-cloud collaborative reinforcement learning optimization, continuously optimize the control strategy through multi-objective reinforcement learning, specifically, the edge RTU collects the state data of the pumping unit and executes the control action, and uploads the state transition record to the cloud, the cloud optimizes the policy model through the deep deterministic policy gradient algorithm, and then issues it to the edge RTU for deployment, forming a closed-loop iteration of data collection, policy optimization, and control execution, wherein the state vector of the model includes pump fullness, current balance, electricity price, and electricity price period; the action vector includes stroke, balance, and start-stop, wherein stroke: adjustment range is -50% to +30%; balance: adjustment range of beam balance block is -10 cm to +10 cm; start-stop: value is 0 or 1, 0 means the pumping unit motor is stopped, and 1 means the pumping unit motor is running; Optimize the policy in the direction of high yield, low consumption, balance, and valley electricity use through weighted reward items, specifically: Yield reward: the higher the ratio of actual liquid production to target liquid production, the higher the reward, which is used to encourage stable production; Energy consumption penalty: the higher the ratio of actual energy consumption to rated energy consumption, the greater the penalty, which is used to suppress high energy consumption; Balance penalty: the greater the deviation of balance from 100%, the greater the penalty, which is used to promote motor load balance; Valley electricity reward: the higher the proportion of valley period production time, the higher the reward, which is used to increase production by using low-price electricity; Dynamic weight adjustment: the cloud updates the weights daily according to the oilfield production target, such as increasing the yield weight in the yield season and increasing the energy consumption weight in the energy-saving season.

[0026] S6: dynamic evaluation and strategy correction, used to continuously track the optimization effect of the strategy and adaptively adjust, specifically including the following steps: S601: evaluation index, specifically power saving rate: compare the total energy consumption of the pumping unit before and after optimization, and automatically calculate and synchronize to the cloud database daily; Liquid production fluctuation monitoring: the deviation of daily average liquid production for 3 consecutive days from the historical average, the liquid production fluctuation refers to the absolute deviation percentage of the daily average liquid production for 3 consecutive days compared with the historical average of the same working day type; Current balance degree Compliance rate, daily statistics of current balance degree The proportion of time within the range of 80%-110%, target value > 90%; S602: Strategy correction, when the liquid production fluctuation monitored for 3 consecutive days is > 5%, collect the pump status data of the past 7 days, including pump fullness, balance degree, electricity price period, and mark the abnormal period, including dynamic liquid surface sudden drop, fault alarm period; based on the abnormal data, recalculate the state transition probability to identify the strategy failure scenario; use the Bayesian optimization algorithm to adjust the learning rate and discount factor parameters of the deep deterministic policy gradient algorithm model to improve the adaptability of the model to abnormal working conditions; If the power saving rate has not been improved for 15 consecutive days and is < 25%, trigger the cloud model retraining, and focus on optimizing the time-of-use electricity price response strategy, such as adjusting the pump stroke increase threshold in the valley period; When the liquid production fluctuation is > 5% for 3 consecutive days When the balance degree compliance rate is < 80%, analyze The working condition characteristics of the period that does not meet the standard, including peak and valley period distribution, pump fullness interval, when the high stroke in the valley period Frequent over-limiting, it is determined that the coordination strategy of stroke adjustment and beam balance block adjustment fails; when retraining the deep deterministic policy gradient algorithm in the cloud, increase the weight of the 5% and related reward items to strengthen the balance degree optimization goal; adjust the trigger threshold of the beam balance block adjustment, for example, from adjust to to improve the adjustment sensitivity; optimize the coordination logic of stroke adjustment and beam balance block adjustment, when the stroke increases by more than 20%, automatically trigger the beam balance block pre-adjustment to compensate for load changes in advance; increase the priority of beam balance block adjustment when The compliance rate is < 70% and the liquid production fluctuation is < 5%, preferentially execute the optimized strategy correction; The cloud model retraining specifically includes data processing: use the labeled abnormal data to expand the training set to ensure that the proportion of abnormal samples is not less than 20%; training strategy: use incremental learning method, retain 70% of the pre-training parameters of the original model, and only fine-tune the network layers related to abnormal scenarios; computing resources: cloud server cluster 16-core CPU + 4 GPU parallel training, single training time < 4 hours; The threshold is derived from a large amount of field experience.

[0027] Example two: A time-sharing optimization design system for pumping wells based on edge computing, comprising an edge perception and data acquisition module, a working condition diagnosis and feature extraction module, a time-sharing electricity price response and prediction module, an edge-end real-time control execution module, an edge-cloud collaborative reinforcement learning module, a safety and compliance assurance module, and a dynamic evaluation and strategy correction module. The edge perception and data acquisition module is electrically connected to the working condition diagnosis and feature extraction module, the working condition diagnosis and feature extraction module is electrically connected to the time-sharing electricity price response and prediction module, the time-sharing electricity price response and prediction module is electrically connected to the edge-end real-time control execution module, the edge-end real-time control execution module is electrically connected to the edge-cloud collaborative reinforcement learning module, the edge-cloud collaborative reinforcement learning module is electrically connected to the dynamic evaluation and strategy correction module, the dynamic evaluation and strategy correction module is electrically connected to the time-sharing electricity price response and prediction module and the edge-end real-time control execution module; The edge sensing and data acquisition module is used to collect the operating parameters of the pumping well in real time and provide real-time operating data with a sampling frequency of 5-10 seconds per time and an accuracy of ±0.5%FS; The operating condition diagnosis and feature extraction module, based on data collected by the edge sensing and data acquisition module, outputs the pumping unit's pump fullness, current balance, and fault type. This module is used to quantitatively assess the well's operating status and identify pumping unit faults such as insufficient fluid supply and valve leakage, providing a basis for production strategy. The time-of-use electricity price response and prediction module divides the time period into peak, flat and valley based on the electricity price. At the same time, the edge RTU synchronizes the grid electricity price in real time and predicts the next 2 hours based on the improved grey prediction model. trends; The edge-side real-time control execution module generates minute-level control instructions based on strategies to adjust the operating parameters of the pumping unit. These strategies include stroke optimization: adjusting the motor speed of the pumping unit through a frequency converter, increasing the stroke rate to 130% of the rated value during off-peak hours; current balance closed-loop: adjusting the position of the walking beam balance block through a stepper motor, with a response time of ≤1 minute; and intermittent pumping cycle adaptation: dynamically adjusting the pumping duration and downtime based on the fluid supply rate. The edge-cloud collaborative reinforcement learning module realizes autonomous evolution of strategies through data-driven, reducing comprehensive energy consumption by 25-42%. Its core framework includes 、 The state vector is composed of the power price and the power price period; the action vector is composed of the pulse frequency adjustment, the adjustment of the walking beam balance block and the start and stop of the pumping unit motor, among which the start and stop are separate strategy branches; Cloud-based deep deterministic policy gradient algorithm optimization strategy, with reward function coupled with output, energy consumption, balance, and valley power ratio; Transfer learning framework, reducing cold start time of new wells by 70%; The security and compliance guarantee module is used for guaranteeing system operation security and data transmission compliance, including edge video analysis, data encryption and zero trust network, wherein the edge video analysis is based on a wide-angle infrared camera+YOLOv5s model to detect illegal behaviors; the data encryption is based on a national secret SM4 algorithm encryption and an HMAC-SHA256 integrity check; the zero trust network is authenticated by bidirectional certificates of edge nodes to block illegal access; The dynamic evaluation and strategy correction module corrects the strategy based on the evaluation indexes of power saving rate, liquid production fluctuation monitoring and balance degree standard rate statistics, so as to ensure that the pumping unit maintains high power saving rate and production stability for a long time.

[0028] The above shows and describes the basic principles and main features of the present application and the advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above-mentioned embodiments, and the above-mentioned embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A time-sharing optimization design method for pumping wells based on edge computing, characterized by: The following steps are included: S1: Edge perception layer construction and real-time data collection, used to build a high-precision, low-latency data collection network covering key operating parameters of pumping wells; S2: Edge layer data extraction and working condition diagnosis. Based on the real-time data collected by S1, it calculates the pumping efficiency index of the pumping unit and diagnoses faults, providing a basis for strategy formulation. S3: Time-of-use electricity price response and operating condition forecast, combined with electricity price period and pump fill degree forecast, to generate a production strategy prototype; S4: Generate real-time control strategies at the edge, dynamically adjusting the pumping frequency based on the pump's fullness and electricity price period, to increase production during off-peak periods and reduce consumption during peak periods. S5: Edge-cloud collaborative reinforcement learning optimization, continuously optimizing control strategies through multi-objective reinforcement learning; S6: Dynamic evaluation and strategy correction, used to continuously track the optimization effect of the strategy and make adaptive adjustments.

2. The time-sharing optimization design method for a pumping well based on edge computing according to claim 1 is characterized in that: In S1, data is collected using a three-phase current sensor, a load displacement sensor, and a pressure sensor. The three-phase current sensor is installed at the three-phase inlet terminal of the pumping unit's motor, collecting the effective value, peak value, and phase difference of the current at a frequency of 5-10 seconds per stroke, and synchronously calculating the pumping unit's motor power. The load displacement sensor is installed axially on the load-bearing rod of the suspension rope at the wellhead of the pumping unit, and is used to monitor the suspension point load and polished rod displacement in real time. The pressure sensor is embedded through a threaded interface at the bottom of the oil pipe of the pumping unit, and is used to monitor changes in the dynamic liquid level depth and calculate the dynamic liquid level change rate based on the depth. The above data is pre-processed by the edge computing node of the edge RTU and transmitted to the edge node. The edge RTU also deploys a CNN model.

3. The time-sharing optimization design method for a pumping well based on edge computing according to claim 1 is characterized in that: Said S2 specifically includes the following steps: S201: Calculate the pump fullness of the pumping unit. Use sliding average filtering to remove high-frequency noise from the load displacement sensor. Use timestamp alignment technology to synchronize the load displacement data with the pumping unit motor current data to the same stroke cycle. Calculate the theoretical stroke area based on the formula: ； in is the theoretical stroke area; is the cross-sectional area of ​​the plunger; Set the stroke length for the pumping unit; The contour detection algorithm based on Canny edge detection is used to identify the closed curve of the dynamometer diagram, and the area within the closed curve is integrated by trapezoidal integration to calculate the effective area of ​​the actual discharge stroke. , in order to extract the actual drainage area, based on which the fullness is quantified: ； S202: Perform peak detection on the three-phase current sensor data, and take the maximum value of the three-phase currents A, B, and C collected by the three-phase current sensor as the upper and lower stroke peak values, which are recorded as and , the sampling period is synchronized with the stroke, and each stroke records a set of peak data, based on which the current balance is calculated: ； when , then it is in equilibrium state; S203: Lightweight fault diagnosis is implemented, using a CNN model deployed on the edge RTU to identify and warn of pumping unit fault types. The CNN model is used to output the pumping unit fault type and its confidence level. Fault types include normal, insufficient fluid supply, valve leakage, and gas lock. When the confidence level of a single fault type is greater than 90% and more than one fault type is output simultaneously, an alarm is triggered.

4. The time-sharing optimization design method for a pumping well based on edge computing according to claim 3 is characterized in that: In the above S201, when The fluid supply to the oil pump is normal; when When the oil pump needs to be supplied with liquid, an early warning is issued; when The oil pump is judged to be insufficiently supplied with fluid.

5. The time-sharing optimization design method for pumping wells based on edge computing according to claim 4 is characterized in that: In S203, the network architecture of the CNN model includes: Input layer: 256×256 pixel grayscale dynamometer diagram; Output layer: Output fault probability types and probability distributions for normal, insufficient fluid supply, valve leakage, and gas lock; The training data set consists of 100,000 labeled dynamometer diagrams, including 40,000 normal working conditions. 、 ; Insufficient liquid supply: 30,000 sheets, i.e. ; Valve loss: 20,000 frames, the power diagram will show sawtooth fluctuations; Gas lock: 10,000 frames, the power diagram will show pear-shaped distortion; Next, the edge RTU automatically scales each dynamometer diagram it collects to 256×256 pixels, performs convolution operations through the FPGA acceleration unit, and outputs the pumping unit's fault type and confidence level. Every time 1,000 new dynamometer diagrams of unlabeled samples are diagnosed, model fine-tuning is automatically started. When the diagnosis confidence is less than 70% for five consecutive times, incremental learning of the CNN model is forced to be triggered.

6. The time-sharing optimization design method for pumping wells based on edge computing according to claim 2 is characterized in that: The S3 specifically includes the following steps: S301: Electricity prices are divided into three time periods: 08:00-12:00 (peak price of 1.2 yuan / kWh), 12:00-18:00 (average price of 0.7 yuan / kWh), and 00:00-06:00 (valley price of 0.3 yuan / kWh). The edge RTU synchronizes the grid electricity price list in real time via 4G, and the local cache is valid for 7 days. S302: Pump fill prediction model implementation, based on the improved grey prediction model: ； The prediction process is: Accumulate and generate ; Solved by least squares method 、 ;Introduction , compensate for the sudden change error of the dynamic liquid level, and predict the next 2 hours When the rate of change of the dynamic liquid level is greater than 10m / h, the power index is automatically adjusted: ； in is the historical pump fullness sequence; 、 is the grey model coefficient, where is the development coefficient, is the ash action amount; is the power correction coefficient; is the power index; To generate a cumulative sequence, Accumulate and generate; is the time step; is a natural constant.

7. The time-sharing optimization design method for pumping wells based on edge computing according to claim 6 is characterized in that: The S4 specifically includes the following steps: S401: Policy triggering conditions include: Increased production during valley period: When the electricity price is low and the pump is full When the rate is higher than 80%, the edge RTU starts the impulse increase mechanism; Peak period consumption reduction: Peak electricity price period and When the speed is less than 50%, the pumping unit motor is switched to one of the low-speed mode of 1 time / minute and shutdown; S402: The implementation process specifically includes: The stroke rate adjustment range is 6 strokes / minute as the rated stroke rate base value, which can be increased to 130% of the rated value during off-peak hours. The low-speed mode is set to 1 stroke / minute, and the pumping unit motor coasts to a stop when shut down. PID control: Dynamically adjust the frequency of the pumping unit's inverter through the PID controller; Real-time feedback: The pumping unit motor current and speed data are collected at each stroke. If the stroke frequency fluctuation exceeds the set value by ±5%, the edge RTU automatically fine-tunes the inverter frequency in 0.5Hz increments. S403: Current balance closed-loop control strategy: prepare a walking beam balance block and install it at the tail of the walking beam of the pumping unit. The walking beam balance block uses a stepper motor to drive a screw transmission mechanism to achieve horizontal position adjustment. Specifically: Real-time monitoring of current balance , automatically adjust the position of the walking beam balance block to balance the current Maintained within the range of 80%-110%, when: If it is less than 80%, it is judged that the downstroke current is too large and the beam balance weight is insufficient; If it is greater than 110%, it is judged that the upstroke current is too large and the balance weight is too heavy. The edge RTU starts the adjustment mechanism: The edge RTU output signal drives the stepper motor, which drives the walking beam balance block to move horizontally along the tail of the walking beam through the screw transmission; The adjustment logic is: When the value is greater than 110%, the beam balance block moves away from the wellhead, increasing the upstroke counterweight and reducing the upstroke current; when When the value is less than 80, the beam balance block moves toward the wellhead, reducing the downstroke load and increasing the downstroke current; After each adjustment, the current peak value of 3 strokes is continuously monitored. If the target range is not reached, a secondary adjustment will be initiated after 1 minute, and a manual intervention alarm will be triggered after a maximum of 3 attempts. S404: Adaptive strategy for intermittent pumping cycle, calculation of fullness The rate of decrease every ten minutes, when the rate of decrease is greater than 0.5% / minute, it is judged that the fluid supply is sufficient and the oil pumping is extended to =40%; when the drop rate is less than 0.2% / min: it is judged that the liquid supply is insufficient and the machine is shut down in advance; S405: Dynamic correction: During shutdown, the pressure sensor is woken up every 30 minutes to monitor the recovery of the dynamic liquid level. If the recovery rate is greater than 0.5% / minute, the shutdown can be ended early. If there is no change in the dynamic liquid level after two consecutive monitorings, the shutdown time will be automatically extended by 20% to avoid frequent starts and stops.

8. The time-sharing optimization design method for pumping wells based on edge computing according to claim 1 is characterized in that: S5 specifically involves the edge RTU collecting pumping unit status data and executing control actions, while also uploading state transition records to the cloud. The cloud optimizes the policy model using a deep deterministic policy gradient algorithm and then sends it to the edge RTU for deployment, forming a closed-loop iteration of data collection, policy optimization, and control execution. The model's state vector includes pump fullness, current balance, electricity price, and electricity price period; the action vector includes stroke rate, balance, and start / stop. The stroke rate has an adjustment range of -50% to +30%. Balance: The adjustment range of the walking beam balance weight is: -10 cm to +10 cm; Start / Stop: The value range is 0 or 1. 0 means the pumping motor is stopped; 1 means the pumping motor is running. The weighted reward items are used to guide the strategy towards high yield, low consumption, balance, and multi-use of off-peak electricity. Specifically: Production Reward: The higher the ratio of actual liquid production to target liquid production, the higher the reward, which is used to encourage the maintenance of stable production; Energy consumption penalty: The higher the ratio of actual energy consumption to rated energy consumption, the greater the penalty, which is used to suppress high energy consumption operations; Balance penalty: The greater the deviation of the balance from 100%, the greater the penalty, which is used to promote motor load balancing; Off-peak electricity reward: The higher the proportion of production time during off-peak hours, the higher the reward, which is used to increase production by utilizing low-priced electricity; Dynamic weight adjustment: The cloud updates weights daily based on oil field production targets.

9. The time-sharing optimization design method for pumping wells based on edge computing according to claim 1 is characterized in that: The S6 specifically includes the following steps: S601: Evaluation indicators, specifically power saving rate: compare the total energy consumption of the pumping units before and after optimization, automatically calculate and synchronize to the cloud database daily; Liquid production fluctuation monitoring: deviation of the daily average liquid production for three consecutive days from the historical average; Current balance Compliance rate, daily statistics of current balance The proportion of time in the range of 80%-110%, with a target value of >90%; S602: Strategy modification: When the monitored fluid production fluctuations exceed 5% for three consecutive days, collect the pumping unit status data for the past seven days and mark the abnormal periods, including periods of sudden drops in the dynamic liquid level and fault alarms. Recalculate the state transition probability based on the abnormal data to identify strategy failure scenarios. Use the Bayesian optimization algorithm to adjust the learning rate and discount factor parameters of the deep deterministic policy gradient algorithm model to improve the model's adaptability to abnormal operating conditions. If the energy saving rate does not increase and is less than 25% for 15 consecutive days, the cloud model will be retrained, focusing on optimizing the time-of-use electricity price response strategy; When 3 consecutive days When the balance rate is less than 80%, analyze The working condition characteristics of the non-standard period include the peak and valley period distribution, pump fullness interval, and when the valley period has a high number of strokes If the limit is exceeded frequently, it is considered that the coordinated strategy of impulse regulation and beam balance block regulation has failed; when retraining the deep deterministic policy gradient algorithm on the cloud, an increase of 5% and The weight of related reward items; optimize the coordinated logic of stroke adjustment and beam balance block adjustment. When the stroke rate increases by more than 20%, the beam balance block pre-adjustment is automatically triggered to compensate for load changes in advance; increase the priority of beam balance block adjustment. When the compliance rate is less than 70% and the liquid production fluctuation is less than 5%, priority will be given to the implementation Optimized strategy revisions.

10. A time-sharing optimization design system for pumping wells based on edge computing, employing the time-sharing optimization design method for pumping wells based on edge computing as described in any one of claims 1 to 9, comprising an edge perception and data acquisition module, a working condition diagnosis and feature extraction module, a time-sharing electricity price response and prediction module, an edge-end real-time control execution module, an edge-cloud collaborative reinforcement learning module, a security and compliance assurance module, and a dynamic evaluation and strategy correction module, characterized in that: The edge sensing and data acquisition module is used to collect the operating parameters of the pumping well in real time and provide real-time operating data with a sampling frequency of 5-10 seconds per time and an accuracy of ±0.5% FS; The working condition diagnosis and feature extraction module outputs the pump fullness, current balance and fault type of the pumping unit based on the data collected by the edge sensing and data acquisition module. This is used to quantitatively evaluate the operating status of the oil well and identify pumping unit faults such as insufficient fluid supply and valve leakage, providing a basis for production strategy. The time-of-use electricity price response and prediction module divides the time period into peak, flat and valley based on the electricity price. At the same time, the edge RTU synchronizes the grid electricity price in real time and predicts the next 2 hours based on the improved grey prediction model. trends; The edge-side real-time control execution module generates minute-level control instructions based on strategies to adjust the operating parameters of the pumping unit. These strategies include stroke optimization: adjusting the motor speed of the pumping unit through a frequency converter, increasing the stroke rate to 130% of the rated value during off-peak hours; current balance closed-loop: adjusting the position of the walking beam balance block through a stepper motor; and intermittent pumping cycle adaptation: dynamically adjusting the pumping duration and downtime based on the fluid supply rate. The edge-cloud collaborative reinforcement learning module realizes the autonomous evolution of strategies through data-driven, reducing the overall energy consumption by 25-42%. Its core framework includes 、 , electricity price, and electricity price period; the action vector is composed of the stroke frequency adjustment, the adjustment of the walking beam balance block, and the start and stop of the pumping unit motor; Cloud-based deep deterministic policy gradient algorithm optimization strategy, with reward function coupled with output, energy consumption, balance, and valley power ratio; Transfer learning framework, reducing cold start time of new wells by 70%; The security and compliance assurance module is used to ensure system operation security and data transmission compliance, including edge video analysis, data encryption and zero-trust network. Edge video analysis is based on a wide-angle infrared camera + YOLOv5s model to detect violations; data encryption is based on the national secret SM4 algorithm encryption, and HMAC-SHA256 integrity verification; Zero-trust network blocks illegal access through two-way certificate authentication at the edge nodes; The dynamic evaluation and strategy correction module corrects the strategy based on the power saving rate, liquid production fluctuation monitoring, and balance compliance rate statistics as evaluation indicators to ensure that the pumping unit maintains a high power saving rate and production stability for a long time.

Images