Rodless oil production well condition parameter prediction method based on ground electric power signal analysis
By analyzing uphole power signals and using CNN-LSTM models, non-intrusive and low-cost online monitoring of well conditions in rodless oil production has been achieved. This solves the problems of high cost and poor reliability of traditional monitoring equipment, and provides real-time and accurate downhole operating parameters, thus providing data support for oilfield management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-26
- Publication Date
- 2026-04-07
AI Technical Summary
Existing downhole monitoring equipment for oil wells is expensive and unreliable. In particular, downhole physical sensors have short service life and require frequent maintenance in harsh environments, and long-distance signal transmission is susceptible to interference, leading to instability in the monitoring system.
By collecting surface power signals and using a CNN-LSTM hybrid prediction model to predict well condition parameters, non-invasive monitoring can be achieved without deploying physical sensors downhole.
It reduces the initial investment cost per well, solves the problems of frequent hardware maintenance and signal interference in traditional monitoring, realizes real-time and accurate monitoring of downhole operating parameters, and improves system reliability and economy.
Smart Images
Figure CN121808262A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of oil extraction equipment monitoring technology, and in particular to a method for predicting rodless oil well condition parameters based on wellhead electrical signal analysis. Background Technology
[0002] In oil extraction operations, obtaining downhole operating parameters such as downhole pressure, downhole temperature, and dynamic fluid level in oil wells in real time and accurately is a core element in optimizing oil production processes and ensuring the safe operation of equipment.
[0003] Currently, the industry commonly uses physical sensors installed directly downhole, such as pressure transmitters, temperature sensors, and level gauges. However, this traditional monitoring technology has the following significant drawbacks: First, in terms of cost, downhole physical sensors are expensive, and multiple types of physical sensors need to be deployed in a single well, resulting in huge initial hardware investment costs. Simultaneously, the extreme and harsh conditions downhole, such as high temperature, high pressure, and corrosion, significantly shorten the lifespan of physical sensors, leading to frequent maintenance and replacements and incurring continuous and high costs, severely restricting the economic viability of oilfield production. Furthermore, regarding system reliability, physical sensor signals need to be transmitted via long-distance cables, making them highly susceptible to interference from the complex electromagnetic environment and mechanical vibrations at the well site, leading to signal attenuation, distortion, or even interruption, making it difficult to guarantee the long-term stability and data reliability of the monitoring system.
[0004] Therefore, there is an urgent need for an innovative well condition prediction method that can solve the problems of high deployment cost and difficult maintenance in existing technologies without relying on downhole physical sensors, and achieve highly reliable and low-cost online monitoring. Summary of the Invention
[0005] The purpose of this invention is to provide a rodless oil well condition parameter prediction method based on wellhead power signal analysis, which overcomes the shortcomings of traditional monitoring technologies. By adopting a non-invasive technical approach, it avoids the deployment of physical sensors downhole, thereby fundamentally solving the technical problems faced by existing technologies, such as high deployment and maintenance costs and poor system reliability due to long-distance signal transmission. This provides an innovative solution for achieving low-cost, high-reliability online monitoring of downhole operating parameters in oilfields.
[0006] To achieve the above objectives, this invention provides a method for predicting rodless oil well condition parameters based on wellhead electrical signal analysis, comprising the following steps: Step S1: Synchronously acquire historical surface power signals and downhole operating parameters of the pumping unit motor; Step S2: Preprocess the historical surface power signal to obtain the preprocessed signal; Step S3: Extract multi-dimensional features from the preprocessed signal and fuse them to construct a multi-dimensional feature vector; Step S4: Construct and train a CNN-LSTM hybrid prediction model; Step S5: Acquire real-time surface power signals, process them through steps S2 and S3 to generate real-time multidimensional feature vectors, input them into the trained CNN-LSTM hybrid prediction model, and output the predicted values of the downhole operating parameters.
[0007] Preferably, in step S1, the wellhead power signal includes a three-phase current signal, a three-phase voltage signal, a DC bus current signal, and a DC bus voltage signal.
[0008] Preferably, in step S1, the downhole operating parameters include downhole pressure, downhole temperature, and oil well dynamic fluid level.
[0009] Preferably, in step S2, the preprocessing includes digitization conversion and bandpass filtering.
[0010] Preferably, step S2 specifically includes: use The sampling rate is used to digitize the historical wellhead power signals to generate corresponding time-domain sequences; Based on the characteristics of historical wellhead power signals, bandpass filtering was applied to the time-domain sequences to obtain preprocessed signals.
[0011] Preferably, the specific expression for bandpass filtering is: ; in, Indicates the current time The input sample value; Indicates the current time The output sampled value; Indicates the previous moment The input sample value; Indicates the first two moments The input sample value; Indicates the previous moment The output sampled value; Indicates the first two moments The output sampled value; , , , , These represent the coefficients of a digital bandpass filter.
[0012] Preferably, step S3 specifically includes: Extract the corresponding feature information from the preprocessed signals; By fusing the feature information of all signals, a multidimensional feature vector is constructed.
[0013] Preferably, the feature information includes frequency domain features and time-frequency domain features.
[0014] Preferably, in step S4, the CNN-LSTM hybrid prediction model includes a convolutional structure and a recurrent structure; wherein, the convolutional structure is used to extract local spatial pattern features of multi-source power signals, and the recurrent structure is used to capture long-term temporal dynamic correlations across signals.
[0015] Therefore, the present invention employs the above-mentioned method for predicting rodless oil well condition parameters based on wellhead electrical signal analysis, and the beneficial technical effects are as follows: (1) This invention uses the power signal from the wellhead that drives the pumping unit motor to perform parameter inversion prediction, without the need to deploy any physical sensors downhole. This not only significantly reduces the initial investment cost of a single well, but also fundamentally solves the industry problem of short lifespan and frequent maintenance of physical sensors in harsh downhole environments. At the same time, it overcomes the inherent defect of poor system reliability caused by interference during long-distance signal transmission.
[0016] (2) This invention has created a new paradigm for oil well monitoring, realizing non-invasive, online real-time monitoring of downhole operating parameters such as downhole pressure, downhole temperature and oil well dynamic fluid level. It transforms the traditional "direct downhole measurement" into "uphole software analysis", realizing a technological leap from hardware dependence to algorithm-driven.
[0017] (3) This invention provides core support for digital management of oilfields. The real-time and accurate downhole operating parameters output provide a reliable data foundation for optimizing oil production processes, implementing predictive maintenance and improving the overall operating efficiency of oilfields. Attached Figure Description
[0018] Figure 1 This is a schematic diagram illustrating the principle of a rodless oil well condition parameter prediction method based on wellhead power signal analysis according to the present invention. Detailed Implementation
[0019] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0020] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.
[0021] Example 1 like Figure 1 As shown, a method for predicting rodless oil well condition parameters based on wellhead electrical signal analysis includes the following steps: Step S1: Synchronously acquire historical surface power signals and downhole operating parameters of the pumping unit motor.
[0022] The surface power signals include three-phase current signals, three-phase voltage signals, DC bus current signals, and DC bus voltage signals. Downhole operating parameters include downhole pressure, downhole temperature, and well dynamic fluid level.
[0023] To achieve synchronous acquisition of historical surface power signals and actual downhole operating parameters, the following method is adopted: First, in the acquisition of surface power signals, a multi-channel synchronous data acquisition card is used to synchronously acquire the three-phase current signal, three-phase voltage signal, DC bus current signal, and DC bus voltage signal driving the pumping unit motor. A high-precision timestamp is added to each batch of acquired surface power signal data frames, which is the historical surface power signal.
[0024] Secondly, in the data acquisition stage of downhole physical sensors (including pressure sensors, temperature sensors, and dedicated sensors for dynamic fluid level (such as echo sounders)), periodic data acquisition is performed through these sensors. When the RTU (Remote Terminal Unit) or host computer receives data points from each downhole physical sensor, it immediately adds a local timestamp to it, which is the true value of the downhole operating parameters.
[0025] To ensure time consistency, the clock of the RTU or host computer must be highly synchronized with the clock of the wellhead power signal acquisition system, thereby ensuring that the two types of data are strictly aligned in time scale and providing accurate aligned supervision data for subsequent model training.
[0026] Step S2: Preprocess the historical wellhead power signals to obtain the preprocessed signals.
[0027] Preprocessing includes digitization and bandpass filtering. Specifically: use The sampling rate is used to digitize the historical wellhead power signals. Specifically, the historical wellhead power signals are converted into voltage signals within the measurable range by a current transformer, and then analog-to-digital conversion is performed by a multi-channel synchronous sampling analog-to-digital converter (ADC) to finally generate the corresponding time-domain sequence. This time-domain sequence is a discrete-time digital signal, which refers to a set of discrete digital sample values arranged at equal time intervals, reflecting the current or voltage value at a certain time.
[0028] Based on the characteristics of historical wellhead power signals, bandpass filtering was applied to the time-domain sequences to obtain preprocessed signals. Specifically: Based on the characteristics of historical wellhead power signals, bandpass filtering is applied to suppress noise and interference in specific frequency bands while retaining effective signal components related to the operating conditions of the pumping unit motor. Bandpass filtering is implemented using a digital bandpass filter, and its operation is described by the following difference equation: ; in, Indicates the current time The input sample value; Indicates the current time The output sampled value; Indicates the previous moment The input sample value; Indicates the first two moments The input sample value; Indicates the previous moment The output sampled value; Indicates the first two moments The output sampled value; , , , , These represent the coefficients of a digital bandpass filter.
[0029] , , , , The parameters are independently set and optimized based on the type of historical wellhead power signal being processed (three-phase current signal, three-phase voltage signal, DC bus current signal, and DC bus voltage signal) and its spectral characteristics (frequency bands containing effective operating condition information and frequency bands with major interference noise). For example, for three-phase current signals and three-phase voltage signals, the coefficient optimization objective is to preserve the fundamental frequency (…). ) and its main harmonics (such as ) components; for DC bus current and DC bus voltage signals, the coefficient optimization objective is to preserve the load dynamic components (such as... It also suppresses high-frequency switching ripple.
[0030] Step S3: Extract multi-dimensional features from the preprocessed signal and fuse them to construct a multi-dimensional feature vector. Specifically: The corresponding feature information is extracted from the preprocessed signal, including frequency domain features and time-frequency domain features.
[0031] Frequency domain characteristics include the amplitude and phase of the fundamental component, the amplitude of specific harmonic components (reflecting load nonlinearity and the influence of power electronic devices), the total harmonic distortion rate (characterizing the degree of waveform distortion), the centroid of the spectrum (characterizing the average position of the spectrum energy distribution), and the bandwidth of the spectrum (characterizing the degree of dispersion of the spectrum around the centroid).
[0032] Time-frequency domain characteristics include signal RMS value, signal peak value and peak-to-peak value, signal average value and absolute average value, waveform factor, peak factor, signal skewness (characterizing the asymmetry of signal distribution), and signal kurtosis (characterizing the sharpness / flatness of signal distribution).
[0033] To eliminate the impact of differences in dimensions and orders of magnitude between different features on model training, all extracted feature information is normalized. Subsequently, all normalized feature information is fused and concatenated in dimensional order to form a one-dimensional high-dimensional feature vector, i.e., the final multi-dimensional feature vector, which serves as the input to the subsequent model.
[0034] Step S4: Construct and train a CNN-LSTM hybrid prediction model. CNN stands for Convolutional Neural Network, and LSTM stands for Long Short-Term Memory Network.
[0035] The CNN-LSTM hybrid prediction model includes convolutional and recurrent structures. It extracts local spatial pattern features of uphole power signals through convolutional structures, and combines this with long short-term memory networks in the recurrent structure to model and capture long-term temporal dynamic correlations across signals. Based on deeply fused multidimensional feature vectors, it outputs real-time estimates of downhole operating parameters.
[0036] The convolutional structure employs two stacked one-dimensional convolutional layers, with the kernel size preferably ranging from 64 to 256 sampling points. The first convolutional kernel can be slightly larger (e.g., 128-256 points) to capture the initial pattern within a wider time window; subsequent convolutional kernels can be appropriately smaller (e.g., 64-128 points) to extract more refined local features. The number of kernels in each layer is typically 32-128, increasing progressively to enrich feature representation; for example, the first layer might have 32 kernels, and the second layer 64. The convolution function uses a standard one-dimensional discrete convolution operation. Each convolutional layer is followed by a... ReLU Activation functions are used to enhance the nonlinear expressive power of the model.
[0037] The recurrent structure uses two stacked LSTM layers, with each LSTM layer typically containing between 128 and 512 hidden units.
[0038] The training process of the CNN-LSTM hybrid prediction model is as follows: (1) Obtain the training dataset; the training dataset includes the multidimensional feature vector extracted in step S3, and the real values of the downhole operating condition parameters synchronously acquired from the surface power signal from which the multidimensional feature vector originates. The real values of the downhole operating condition parameters are used as supervision labels for model training. To ensure the generalization ability of the CNN-LSTM hybrid prediction model, the training dataset is divided into mutually exclusive training, validation, and test sets in a preset ratio of 7:2:1.
[0039] (2) Model initialization and optimizer configuration: The weight parameters of the CNN-LSTM hybrid prediction model are randomly initialized. A suitable optimization algorithm (such as Adam or RMSprop) is selected, and hyperparameters such as the initial learning rate and weight decay coefficient are set. At the same time, a loss function suitable for multi-task regression prediction is defined, usually using mean squared error or smoothed average absolute error as the objective function to measure the difference between the model's predicted value and the true label.
[0040] (3) Iterative training and backpropagation: The training set data is input into the CNN-LSTM hybrid prediction model, and forward propagation is performed to generate predicted values of downhole operating parameters based on the current parameters. The loss value between the predicted values of downhole operating parameters in this batch and the corresponding true values of downhole operating parameters is calculated. Then, the backpropagation algorithm is used to calculate the gradient of the loss function with respect to the parameters of each layer of the CNN-LSTM hybrid prediction model. The configured optimizer is called to update the parameters of the CNN-LSTM hybrid prediction model according to the calculated gradient, so as to gradually minimize the loss function.
[0041] (4) Training process monitoring and verification: After each training round, the current model is used to make predictions on the validation set and the validation loss is calculated.
[0042] (5) The training process continues iteratively until any of the following stopping conditions are met: ①Early stopping condition: The validation set loss is within the range of consecutive... N No decline within training rounds ( N (Preset positive integer).
[0043] ② Reaching the maximum number of training rounds: The total number of training rounds exceeds the preset limit. T ( T (is a positive integer).
[0044] After training is terminated, the model parameters corresponding to the lowest loss on the validation set during the entire training process are selected as the final CNN-LSTM hybrid prediction model.
[0045] Step S5: Acquire real-time surface power signals, process them through steps S2 and S3 to generate real-time multidimensional feature vectors, input them into the trained CNN-LSTM hybrid prediction model, and output the predicted values of the downhole operating parameters.
[0046] Therefore, the present invention adopts the above-mentioned rodless oil well condition parameter prediction method based on well power signal analysis, which effectively solves the problems of high deployment cost, difficult maintenance and insufficient reliability of traditional monitoring technology, and realizes non-intrusive and low-cost online monitoring of oilfield downhole operating parameters.
[0047] 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 them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A method for predicting rodless oil well condition parameters based on wellhead electrical signal analysis, characterized in that, Includes the following steps: Step S1: Synchronously acquire historical surface power signals and downhole operating parameters of the pumping unit motor; Step S2: Preprocess the historical surface power signal to obtain the preprocessed signal; Step S3: Extract multi-dimensional features from the preprocessed signal and fuse them to construct a multi-dimensional feature vector; Step S4: Construct and train a CNN-LSTM hybrid prediction model; Step S5: Acquire real-time surface power signals, process them through steps S2 and S3 to generate real-time multidimensional feature vectors, input them into the trained CNN-LSTM hybrid prediction model, and output the predicted values of the downhole operating parameters.
2. The method for predicting rodless oil well condition parameters based on wellhead electrical signal analysis according to claim 1, characterized in that, In step S1, the wellhead power signals include three-phase current signals, three-phase voltage signals, DC bus current signals, and DC bus voltage signals.
3. The method for predicting rodless oil well condition parameters based on wellhead electrical signal analysis according to claim 1, characterized in that, In step S1, the downhole operating parameters include downhole pressure, downhole temperature, and oil well dynamic fluid level.
4. The method for predicting rodless oil well condition parameters based on wellhead electrical signal analysis according to claim 1, characterized in that, In step S2, the preprocessing includes digitization conversion and bandpass filtering.
5. The method for predicting rodless oil well condition parameters based on wellhead electrical signal analysis according to claim 4, characterized in that, Step S2 is as follows: use The sampling rate is used to digitize the historical wellhead power signals to generate corresponding time-domain sequences; Based on the characteristics of historical wellhead power signals, bandpass filtering was applied to the time-domain sequences to obtain preprocessed signals.
6. The method for predicting rodless oil well condition parameters based on wellhead electrical signal analysis according to claim 5, characterized in that, The specific expression for bandpass filtering is: ； in, Indicates the current time The input sample value; Indicates the current time The output sampled value; Indicates the previous moment The input sample value; Indicates the first two moments The input sample value; Indicates the previous moment The output sampled value; Indicates the first two moments The output sampled value; , , , , These represent the coefficients of a digital bandpass filter.
7. The method for predicting rodless oil well condition parameters based on wellhead electrical signal analysis according to claim 1, characterized in that, Step S3 is as follows: Extract the corresponding feature information from the preprocessed signals; By fusing the feature information of all signals, a multidimensional feature vector is constructed.
8. The method for predicting rodless oil well condition parameters based on wellhead electrical signal analysis according to claim 7, characterized in that, The feature information includes frequency domain features and time-frequency domain features.
9. The method for predicting rodless oil well condition parameters based on wellhead electrical signal analysis according to claim 1, characterized in that, In step S4, the CNN-LSTM hybrid prediction model includes convolutional and recurrent structures; the convolutional structure is used to extract local spatial pattern features of multi-source power signals, and the recurrent structure is used to capture long-term temporal dynamic correlations across signals.