A method for predicting oil well production based on PSO-VMD-LSTM

Abstract

This paper proposes a PSO-VMD-LSTM-based oil well production forecasting method. This method builds on the traditional VMD-LSTM forecasting method by applying the PSO algorithm to both the VMD model and the LSTM network. Specifically, the PSO algorithm optimizes the VMD model to obtain the optimal variational mode decomposition (VMD) and penalty factor, avoiding the errors caused by observational VMD and penalty factor calculations. Simultaneously, the PSO algorithm optimizes the LSTM network, enabling optimal parameterization, further improving the accuracy of the forecasting method.

Description

Technical Field

[0001] The present invention relates to the field of oil well production prediction, and in particular to an oil well production prediction method based on PSO-VMD-LSTM Background Art

[0002] In recent years, with the increasing difficulty of oil exploration, oilfield production technology has become increasingly reliant on data analysis and prediction techniques. Oil well yield prediction, as a key technology, is crucial to oilfield development and management. Traditional oil well yield prediction methods rely primarily on empirical or statistical models. However, due to issues such as data quality and missing data, these methods have significant limitations in both accuracy and robustness. Another oil well yield prediction method uses numerical simulation. However, the model assumptions used in numerical simulations may be inaccurate and fail to account for some difficult-to-predict factors in real-world situations. Furthermore, numerical simulations typically require significant computational resources and time, resulting in high costs. Therefore, in the field of oil well yield prediction, there is an urgent need for efficient, accurate, and stable prediction methods.

[0003] Prior art CN115809711A - An MLR oil well production prediction method based on improved EMD. This method addresses the shortcomings of MLR through EMD decomposition and Spearman analysis. This method significantly improves the accuracy of MLR predictions from the perspective of input variable construction and selection. This technology minimizes the raw data dimensions required, reducing the cost of collecting oilfield development indicators. The MLR model primarily predicts past data and may not accurately capture long-term trend changes. Oil well production time series data often exhibits non-stationary characteristics such as trends and periodicity, which makes the MLR model perform poorly when processing oil well production time series data, resulting in inaccurate prediction results.

[0004] Prior art CN 110909931A - A well logging curve prediction method based on modal decomposition and reconstruction and a deep LSTM-RNN model. This method reconstructs CEEMD decomposition items with similar fluctuation frequencies into three new components with typical characteristics: a high-frequency term, a low-frequency term, and a trend term. The reconstructed new components are then normalized and divided into training and test data. A deep LSTM-RNN model is established for each component and trained using the training data. The trained model is then used to predict missing or distorted well logging curves. Finally, the prediction results of the three components are reconstructed and denormalized to obtain the well logging curve prediction results. However, CEEMD requires multiple integrated decompositions, is not suitable for processing large-scale oil well production data, and has insufficient prediction accuracy. Summary of the Invention

[0005] In order to solve the problem of low prediction accuracy of existing oil well production prediction methods, the present invention uses the VMD (Variational Mode Decomposition)-LSTM (Long Short-Term Memory) method to predict oil well production. Based on this method, the PSO (Particle Swarm Optimization) algorithm is used to optimize the VMD model and LSTM network respectively, thereby improving the quality of oil well production prediction.

[0006] Technical solution:

[0007] Step 1: Select an oil well as the experimental object and use its production data for prediction and verification. 80% of the data is used for prediction and 20% for verification. The prediction results obtained using the 80% prediction data are compared with the 20% verification data. If the error is less than the set tolerance, the method is proven to be effective. The production data of the oil well is scaled to the interval [0, 1] using the normalization method. Normalization is to bring all data to the same order of magnitude and eliminate the adverse effects of singular data on the prediction.

[0008] Step 2: Use wavelet transform to reduce noise on normalized data. The wavelet transform formula is:

[0009]

[0010] Step 3: Use the VMD model optimized by the PSO algorithm to obtain the decomposition modal results of the normalized data.

[0011] Working principle of VMD model:

[0012] The process of VMD model processing data can be regarded as solving constrained variation problem. The model is shown in formula (1):

[0013]

[0014]

[0015] In formula (1), t is time, the intrinsic modal components {uk} = {u1, u2, u3, ···, uk}, k is the number of intrinsic modal components, the center frequency of the intrinsic modal components {ωk} = {ω1, ω2, ω3, ···, ωk}, δ(t) is the Dirichlet function, and f is the original signal.

[0016] In order to find the optimal solution of formula (1), a Lagrangian multiplier λ and a penalty factor α are introduced to construct the augmented Lagrangian function of the structure, as shown in formula (2):

[0017]

[0018] By using the alternating direction multiplier algorithm to solve Equation (2), the original signal can be decomposed into k intrinsic modal components.

[0019] The PSO algorithm is used to optimize the above VMD model:

[0020] (1) Set the initial parameter values ​​and optimization function of the particle swarm algorithm and determine the value range of the parameter [k, α];

[0021] (2) Initialize the velocity and position of the particle, and set the parameter [k, α] to the position of the particle;

[0022] (3) Bring the initial parameters [k, α] into the VMD model to decompose the signal and obtain k intrinsic modal components. Calculate the information entropy of each intrinsic modal component, compare the information entropy values ​​of all particles, and update the local minimum of the individual and the global minimum of the population.

[0023] (4) Update the particle's velocity and position to update the parameters [k,α];

[0024] (5) Repeat steps (3)-(4) until the maximum number of iterations is reached, and output the global minimum of the population and the position of the corresponding particle. The global minimum of the population is the minimum entropy value, and the position of the particle is the parameter combination [k, α] corresponding to the entropy value. The parameter combination [k, α] is the optimal number of intrinsic modal components and penalty factor corresponding to this model.

[0025] Step 4: Use the PSO algorithm to optimize the LSTM network and input the intrinsic modal component results of the normalized data into the LSTM network.

[0026] Working Principle of an LSTM Network: Each recurrent unit in an LSTM has a new unit state C, which retains previously input information over a long period of time. The key to LSTM is the control of this long-term state C. The LSTM neural network module utilizes a "gate" structure, including an input gate, a forget gate, and an output gate. The input gate allows information from the previous unit to be input into the long-term state C for long-term retention. The forget gate controls the elements in the previous state Ct-1 that are forgotten. The output gate controls the output of the long-term state C from the LSTM module at time t.

[0027] The PSO algorithm is used to optimize the various parameters of the LSTM network:

[0028] (1) Set the PSO internal algorithm parameters such as the number of particles, search dimension, position and speed of all particles, so as to obtain the best position experienced by the individual and the global best position, and the historical best fitness value of each individual;

[0029] (2) Set the LSTM search parameters such as LSTM time step, number of hidden layer neurons, number of hidden layers, learning rate, etc., take the batch size as the optimization object, and initialize the adaptive PSO algorithm;

[0030] (3) Calculate the initial particle fitness, initial global optimum, and fitness value, thereby updating the initial global optimal parameters, fitness function, and individual optimal parameters;

[0031] (4) Obtaining the optimal parameter combination: training the LSTM network and using the PSO algorithm to find the best global optimal parameters;

[0032] (5) Construct LSTM prediction network: Construct LSTM network using optimal parameters.

[0033] The best global optimal parameters found by the PSO algorithm are the optimal network parameters corresponding to the decomposition modal results. The LSTM network optimized by the PSO algorithm has better prediction accuracy.

[0034] Step 5: The LSTM network outputs the prediction results for each mode, sums the prediction results of each intrinsic modal component and denormalizes them to obtain the final prediction result.

[0035] The root mean square error is used as the criterion for judging the prediction accuracy. The root mean square error formula is:

[0036]

[0037] The beneficial effects of the present invention are as follows: the VMD model is optimized by the PSO algorithm, so that the VMD model no longer needs to use the observation method to estimate the number of intrinsic modal components k and the penalty factor α; the LSTM network is optimized by the PSO algorithm, so that the LSTM network can be set with optimal parameters; the optimized VMD model and LSTM network work together to obtain more accurate prediction capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 : Flow chart of the present invention.

[0039] Figure 2 : PSO algorithm optimizes VMD fitness evolution graph.

[0040] Figure 3 : VMD decomposition modality images.

[0041] Figure 4 : Comparison chart of the prediction results of each method. DETAILED DESCRIPTION

[0042] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0043] Figure 1 This is a specific flow chart of the present invention using the PSO algorithm to optimize the VMD model and LSTM network to predict oil well production.

[0044] The oil well production forecast process is as follows:

[0045] Step 1: Select an oil well as the experimental object, use the production data of the oil well for prediction and verification, use 80% of the data for prediction and 20% of the data for verification, compare the prediction results obtained using 80% of the prediction data with the 20% of the verification data, and if the error is less than the set tolerance value, it proves that the prediction method is effective.

[0046] The production data of the oil well is scaled to the interval [0,1] using the normalization method. Normalization is to make all data at the same order of magnitude, eliminate the adverse effects of singular data, and avoid the large difference between data values ​​affecting the prediction results.

[0047] Step 2: Use wavelet transform to process the normalized data. The wavelet transform formula is:

[0048]

[0049] Where WT(a,τ) is the transformed signal, a represents the expansion and contraction of the wavelet function, τ represents the translation of the control wavelet function, and f(t) represents the signal before the transformation. Wavelet transform can be used to reduce noise in data and improve data quality.

[0050] Step 3: Use the VMD model optimized by the PSO algorithm to obtain the decomposition modal results of the normalized data.

[0051] Working principle of VMD model:

[0052] The process of VMD model processing data can be regarded as solving constrained variation problem. The model is shown in formula (1):

[0053]

[0054]

[0055] In formula (1), t is time, the intrinsic modal component {uk} = {u1,u2,u3,···,uk}, k represents the number of intrinsic modal function components, the center frequency of the intrinsic modal component {ωk} = {ω1,ω2,ω3,···,ωk}, δ(t) is the Dirichlet function, and f is the original signal.

[0056] In order to find the optimal solution of formula (1), a Lagrangian multiplier λ and a penalty factor α are introduced to construct the augmented Lagrangian function of the structure, as shown in formula (2):

[0057]

[0058] By using the alternating direction multiplier algorithm to solve Equation (2), the original signal can be decomposed into k intrinsic modal components.

[0059] The PSO algorithm is used to optimize the above VMD model:

[0060] (1) Set the initial parameter values ​​and optimization function of the particle swarm algorithm and determine the value range of the parameter [k, α];

[0061] (2) Initialize the velocity and position of the particle, and set the parameter [k, α] to the position of the particle;

[0062] (3) Bring the initial parameters [k, α] into the VMD model to decompose the signal and obtain k intrinsic modal components. Calculate the information entropy of each component, compare the information entropy values ​​of all particles, and update the local minimum of the individual and the global minimum of the population.

[0063] (4) Update the velocity and position of the particle to update the parameters [k,α];

[0064] (5) Repeat steps (3)-(4) until the maximum number of iterations is reached, and output the global minimum of the population and the position of the corresponding particle. The global minimum of the population is the minimum entropy value, and the position of the particle is the parameter combination [k, α] corresponding to the entropy value. The parameter combination [k, α] is the optimal number of modal components and penalty factor corresponding to this model.

[0065] In this implementation, the VMD model optimized by the PSO algorithm is obtained with the optimal modal component number k being 5 and the penalty factor α being 2000. The fitness of the VMD model optimized by the PSO algorithm is as follows: Figure 2 As shown, the images of each modal component are as follows Figure 3 shown.

[0066] Step 4: Use the PSO algorithm to optimize the LSTM network and input the decomposition modal results of the normalized data into the LSTM network.

[0067] Working Principle of an LSTM Network: Each recurrent unit in an LSTM has a new unit state C, which stores the previous input information over a long period of time. The key to LSTM is the control of this long-term state C. The LSTM neural network module utilizes a "gate" structure, including an input gate, a forget gate, and an output gate. The input gate allows information from the previous unit to be input into the long-term state C for long-term retention; the forget gate controls the elements in the previous state Ct-1 that are forgotten; and the output gate controls the output of the long-term state C from the LSTM module at time t.

[0068] The PSO algorithm is used to optimize the various parameters of the LSTM neural network model:

[0069] (1) Set the PSO internal algorithm parameters such as the number of particles, search dimension, position and speed of all particles, so as to obtain the best position experienced by the individual and the global best position, and the historical best fitness value of each individual;

[0070] (2) Set the LSTM search parameters such as LSTM time step, number of hidden layer neurons, number of hidden layers, learning rate, etc., take the batch size as the optimization object, and initialize the adaptive PSO algorithm;

[0071] (3) Calculate the initial particle fitness, initial global optimum, and fitness value, thereby updating the initial global optimal parameters, fitness function, and individual optimal parameters;

[0072] (4) Obtain the optimal parameter combination: train the LSTM network model and use the PSO algorithm to find the best global optimal parameters; (5) Construct the LSTM prediction network: construct the LSTM network using the optimal parameters.

[0073] The best global optimal parameters found by the PSO algorithm are the optimal network parameters corresponding to the decomposition modal results. The LSTM network optimized by the PSO algorithm has better prediction accuracy.

[0074] The optimal parameters of the LSTM network obtained in this embodiment are: time step of 6, number of hidden layer neurons of 15, number of hidden layers of 1, learning rate of 0.001, and batch size of 32.

[0075] Step 5: The LSTM network outputs the prediction results for each modality. The prediction results for each modality are summed and denormalized to obtain the final prediction result.

[0076] Using the root mean square error as the criterion for prediction accuracy, the root mean square error formula is:

[0077]

[0078] The smaller the root mean square error (RMSE), the more accurate the prediction result. The root mean square error data of the prediction method used in the present invention and other prediction methods are shown in Table 1.

[0079] Table 1

[0080] Model RMSE LSTM 2.81 VMD-LSTM 1.154 PSO-VMD-LSTM 0.598

[0081] Comparison of oil well production prediction method based on PSO-VMD-LSTM with other prediction methods Figure 4 As shown in the figure, the oil well production prediction method based on PSO-VMD-LSTM is the most accurate.

[0082] Finally, it should be noted that the purpose of disclosing the embodiments is to facilitate a further understanding of the present invention. However, those skilled in the art should understand that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the contents disclosed in the embodiments, and the scope of protection claimed by the present invention shall be determined by the scope defined in the claims.

Claims

1. A method for predicting oil well production based on PSO-VMD-LSTM, characterized in that: The following steps are involved: Step 1: Normalize the oil well production data; Step 2: Use wavelet transform to reduce noise of normalized data; Step 3: Use the VMD model optimized by the PSO algorithm to obtain the decomposition modal results of the normalized data; in the process of optimizing the VMD model by the PSO algorithm, the initial parameters are introduced into the VMD model to decompose the signal to obtain k intrinsic modal components, calculate the information entropy of each component, compare the information entropy values ​​of all particles, and update the local minimum of the individual and the global minimum of the population; Step 4: Use the PSO algorithm to optimize the LSTM network and input the decomposition modal results of the normalized data into the LSTM network. During the PSO algorithm optimization of the LSTM network, set the LSTM search parameters, take the batch size as the optimization object, and initialize the adaptive PSO algorithm. Step 5: The LSTM network outputs the prediction results for each modality. The prediction results for each modality are summed and denormalized to obtain the final prediction result.

2. The oil well production prediction method based on PSO-VMD-LSTM according to claim 1, characterized in that: The process of optimizing the VMD model using the PSO algorithm is as follows: (1) Set the initial parameter values ​​and optimization function of the particle swarm algorithm, and determine the value range of the number of intrinsic modal components k and the penalty factor α; (2) Initialize the particle's velocity and position and determine the initial values ​​of the parameters [k, α]; (3) Bring the initial parameters [k, α] into the VMD model to decompose the signal; (4) Update the inertia weight, learning factor, velocity, position and reconstruct particles according to the formula; (5) Update the particle's velocity and position to update the parameters [k, α]; (6) After reaching the maximum number of iterations, the optimal parameters [k, α] are obtained.

3. The oil well production prediction method based on PSO-VMD-LSTM according to claim 1, characterized in that: 80% of oil well production data is used for prediction and 20% is used for verification.

4. The oil well production prediction method based on PSO-VMD-LSTM according to claim 3 is characterized in that: The comparison between the prediction results and the actual verification data uses the root mean square error as the evaluation criterion.

5. The oil well production prediction method based on PSO-VMD-LSTM according to claim 1, characterized in that: The LSTM network uses a gate algorithm, which includes a forget gate algorithm, an input gate algorithm, and an output gate algorithm.

6. The oil well production prediction method based on PSO-VMD-LSTM according to claim 1 or 5, characterized in that: The LSTM network uses a gating algorithm to control the long-term state of input information.

7. The oil well production prediction method based on PSO-VMD-LSTM according to claim 1 or 5, characterized in that: The process of PSO algorithm optimizing LSTM network is as follows: (1) Set the number of particles, search dimension, and position and velocity parameters of all particles in the PSO algorithm; (2) Set the time step, number of hidden layer neurons, and learning rate parameters of the LSTM network; (3) Calculate the initial particle fitness, initial global optimum, and fitness value parameters; (4) Find the optimal parameter combination of LSTM network through PSO algorithm; (5) Build the optimal LSTM network.

8. The oil well production prediction method based on PSO-VMD-LSTM according to claim 7, characterized in that: The optimal parameter combination of the LSTM network found by the PSO algorithm is the optimal network parameter corresponding to the decomposition modal result.

Images