Oil well production prediction method based on kalman filter under spatio-temporal neural network model

Abstract

A kind of oil well production prediction method based on Kalman filtering under space-time neural network model, the dynamic data of target area well is filtered by Kalman filtering, then the filtered dynamic data is input into space-time neural network model to train model, and the trained model is used to predict the future oil production of target oil well.The present application utilizes the convolution network characteristics to process the time lag of data on one hand, and utilizes the GCN network layer to obtain the spatial relationship between the target area well network on the other hand, utilizes the TCN to analyze the time sequence characteristic relationship, and truly realizes the space-time prediction of oilfield.In addition, in the prediction aspect, the weighted loss function of multiple outputs is optimized and trained by Adam, and the whole prediction model parameters are updated in real time, which improves the prediction accuracy of future oil production of oil well and the practicability of prediction model.

Description

Technical Field

[0001] This invention relates to a technology in the field of well network development, specifically a method for predicting oil well production based on a spatiotemporal neural network model using Kalman filtering. Background Technology

[0002] Water-drive well network development is the most commonly used method in conventional oil reservoir development. Predicting the production of water-drive well networks is the foundation and basis for decision-making issues such as assessing oilfield development efficiency, formulating oilfield development plans, and designing and adjusting well network injection and production implementation schemes. Rapid and accurate production prediction for water-drive reservoir well networks is of great significance for the regulation and optimization of injection and production well network operating procedures in the efficient development of water-drive reservoirs. Summary of the Invention

[0003] This invention addresses the problem that existing technologies suffer from the direct impact of noise in the raw data on prediction results. It proposes an oil well production prediction method based on a spatiotemporal neural network model using Kalman filtering. The method considers the influence of inherent data noise on the prediction results and performs noise reduction processing to improve prediction accuracy. The spatiotemporal neural network model comprises four modules: a data preprocessing module, a spatial relationship acquisition module, a time-series prediction module, and an output module. The data preprocessing module handles time delays and reconstructs the dataset. Considering the heterogeneity of oil reservoir geology and the non-immediacy of water injection effects, the water injection does not immediately cause changes in oil well production due to the reservoir's heterogeneity; there is a certain time delay and dissipation involved. This paper proposes a new dynamic dataset of well networks in the target area. It utilizes a one-dimensional convolutional neural network to process water injection data with time delays and combines this data with filtered oil production data from oil wells. In the spatial relationship acquisition module, spatial relationship analysis is performed using a GCN network layer by incorporating graph data, fully considering the impact of the dynamics of surrounding wells on the production of the oil wells. In the time series prediction module, a temporal convolutional network (TCN) is used to analyze the data from a temporal perspective. The output module utilizes the outputs of multiple single-layer perceptron networks to generate multiple predicted values, namely, the predicted future oil production values ​​of multiple oil wells in the target area. In practical application, a method for updating the prediction model is proposed to increase its practicality.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to an oil well production prediction method based on a spatiotemporal neural network model using Kalman filtering. The method involves filtering the dynamic data of wells in the target area using Kalman filtering, then inputting the filtered dynamic data into the spatiotemporal neural network model for training. The trained model is then used to predict the future oil production of the target oil well and is updated periodically.

[0006] The dynamic data includes: the monthly water injection volume of injection wells and the monthly oil production volume of oil-producing wells in the target area well network.

[0007] The spatiotemporal neural network model includes a data preprocessing module, a spatial relationship acquisition module, a time series prediction module, and an output module. Specifically, the data preprocessing module uses a one-dimensional convolutional neural network to process the filtered historical dynamic data of the injection wells with time delay and outputs the resulting labeled data and historical dynamic data to the spatial relationship acquisition module. The spatial relationship acquisition module extracts the spatial features between injection and production wells in the target area's well network and outputs them to the time series prediction module. The time series prediction module uses a two-layer TCN to fit the spatial features. The output module uses multiple single-layer perceptrons to predict the oil production of multiple oil wells and outputs the predicted oil production of the target oil well based on the fitting results. Furthermore, in terms of prediction, the model is updated over a certain period to improve its practicality.

[0008] The aforementioned time-lag processing refers to: constructing a one-dimensional convolutional network to perform time-lag processing on the monthly water injection volume of injection wells. The convolution kernel size is set to n, generally representing the extraction and compression of temporal features between n months of water injection volume. The convolution direction is from front to back along the time dimension. After convolution, the monthly water injection volume data will be reduced by n-1 time steps in the time dimension. Subsequently, the time-lag processed data is combined with the monthly oil production data of selected oil wells to reconstruct a feature dataset. The first n-... The first step involves a rejection process. The purpose of this step is to remove the oil production data from the previous n-1 time steps, since the oil production at this moment is the cumulative effect of water injection from the previous n-1 time steps, which is a time lag. The water injection and oil production data, which have been processed as described above, are then integrated into a new dataset. At the same time, the prediction dataset is reconstructed. The data from k time steps are used as the feature dataset, and the oil production data from the target oil wells in k+1 time steps are integrated into the label dataset. Finally, the time series dataset required for prediction is obtained.

[0009] The aforementioned extraction of the spatial characteristics between injection and production wells in the target area well network refers to obtaining the spatial relationship between water injection wells and oil production wells in the target area through a single GCN network layer, specifically... Wherein: H (t) For the feature input at time node t, A is the adjacency matrix, representing the graph data of the spatial relationship between well points in the corresponding target area, D is the degree matrix, and the diagonal elements of the diagonal matrix are obtained by adding the elements of each row of A, I N This is the corresponding identity matrix.

[0010] The fitting process described above is achieved through a two-layer TCN.

[0011] The aforementioned graph data refers to: by collecting the geometric information of the target well network, constructing a spatial relationship graph between well points in the injection and production well network based on the topological characteristics of the injection and production wells, and representing it using an adjacency matrix A. Attached Figure Description

[0012] Figure 1 This is a flowchart of the present invention;

[0013] Figure 2 This is a schematic diagram of the basic spatial structure provided in an embodiment of the present invention;

[0014] Figure 3 This is a schematic diagram of the architecture of a spatiotemporal neural network model. Detailed Implementation

[0015] like Figure 1 As shown, this embodiment relates to an oil well production prediction method based on a spatiotemporal neural network model using Kalman filtering, specifically including:

[0016] Step 1) Collect the geometric information of the target well network, construct a graph neural network structure for the injection and production well network based on the topological characteristics of the injection and production wells, and use graph representation to learn and construct the graph data, i.e., the adjacency matrix, which specifically includes:

[0017] 1.1) The relationship between water injection wells and oil production wells is represented graphically, such as... Figure 2 As shown.

[0018] 1.2) Represent the graph data as the following adjacency matrix: [[01010100] [10101010] [01010101] [10101010] [01010101] [10101010] [01010101] [10101010]]

[0027] 1.3) Save the adjacency matrix to the data preprocessing module.

[0028] Step 2) Collect the monthly water injection volume of the injection wells and the monthly oil production of the oil production wells in the target area well network;

[0029] Step 3) Calculate the optimal estimates of monthly water injection volume and monthly oil production volume based on the historical dynamic data of monthly injection wells and production wells using Kalman filtering. This includes:

[0030] 3.1) Calculate the optimal estimate of the monthly water injection volume at time k. Where: z kH represents the actual water injection volume at time k; k The observation matrix represents the degree to which the actual system state depends on the calculated data;

[0031] 3.2) Calculate the Kalman gain K k =P k H k T (H k P k H k T +R k ) -1 Error covariance matrix of monthly water injection volume after Kalman filtering Wherein: H k The observation matrix represents the dependence of the actual system state on the calculated data; in practice, it is set to 1, and R... k For noise covariance, P k K represents the error covariance moment corresponding to the predicted monthly water injection volume at time K. k The Kalman gain is used to iterate through the historical dynamic data of the well network;

[0032] The filtered data is divided into a training set and a test set;

[0033] Step 4) Establish a spatiotemporal neural network model including a data preprocessing module, a spatial relationship acquisition module, a time series prediction module, and an output module. The model uses historical dynamic data of the target area well network processed in Step 3) and the graph data from Step 1) as input, specifically including:

[0034] Step 4.1) Due to the heterogeneity of reservoir geology, water injection does not immediately cause changes in the oil production of the well, i.e., there is a certain time lag and dissipation. This method utilizes the characteristics of one-dimensional convolutional networks to establish a one-layer one-dimensional convolutional neural network to process the monthly water injection volume of the injection well in a time lag manner. The size of the convolution kernel is set to n, which means extracting the time series features between the water injection volumes of n months to consider the lag phenomenon in the propagation process, so that the inversion results are more consistent with the actual situation of the reservoir.

[0035] Step 4.2) Filter the monthly oil production data of the oil wells, remove the monthly oil production data of the first n-1 months of the data list, and combine it with the data set required by the water injection data component prediction unit module after convolution.

[0036] Step 4.3) Perform feature data and label data division on the data group obtained in Step 4.2). Use the data from k time steps as feature data and integrate the oil production data of the target oil well from k+1 time steps as labels.

[0037] Step 4.4) Use the data obtained in Step 4.3), the graph data, as input to the spatial relationship acquisition module, i.e., a single-layer GCN network layer, and use the ReLU function as the activation function;

[0038] Step 4.5) The output-input timing prediction module obtained from step 4.4) uses the ReLU function as the activation function.

[0039] Step 4.6) The output obtained from step 4.5) is the output layer composed of single-layer perceptrons. The output is the predicted monthly oil production of the target oil well. Based on the example, the monthly oil production of 4 oil wells is predicted. Therefore, the outputs of the four single-layer perceptrons correspond to the predicted values ​​of the four oil wells respectively.

[0040] Step 4.7) Optimize the spatiotemporal neural network model by calculating the mean square error (MSE) between the target well's monthly oil production in the training set and the model's predicted monthly oil production. Specifically, the Adam optimization algorithm is used to weight the four loss functions. The weight of each well's loss function is the proportion of that well's production to the total production of the four wells over k time steps in this training process. in: Let be the oil production of the i-th well at time step t. The total loss function is: MSE i These represent the mean square error between the predicted output and the actual value of each of the four wells.

[0041] Step 5) Predict the production output of four oil wells in the target area using the trained spatiotemporal neural network model. The specific steps are as follows:

[0042] Step 5.1) Using the dynamic data of the target well network from the latest k time steps, input the model to predict the monthly oil production of the four oil wells in the future k+1 time step.

[0043] Step 5.2) Based on the predicted oil production results, adjust the water injection plan for step k+1 and execute the water injection according to this plan to obtain the actual oil production of the target area oil wells in step k+1. Then, incorporate the obtained actual value of step k+1 into the historical data.

[0044] Step 5.3) Re-input the updated data from the latest k time steps into the trained model to obtain the predicted oil production of the target area wells in the (k+1)th time step of the new period, and repeat the above steps.

[0045] Step 5.4) Since the reservoir system changes slowly, the model that has been trained based on historical data may need to be readjusted. The prediction system is retrained and adjusted. Every 50 time steps, the latest 500 steps of data are selected and the process after step 35 is re-executed to achieve continuous updating of the entire prediction model.

[0046] Compared with existing technologies, this invention utilizes the characteristics of convolutional networks to process data with time delays, and uses GCN network layers to obtain the spatial relationships between well networks in the target area, and uses TCN to analyze the temporal characteristic relationships, thus achieving true spatiotemporal prediction of oil fields and improving the prediction accuracy of future oil production from oil wells.

[0047] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A method for predicting oil well production based on a spatiotemporal neural network using Kalman filtering, characterized in that, The dynamic data of the target area wells are filtered by Kalman filtering, and then the filtered dynamic data is input into the spatiotemporal neural network model for training. The trained model is then used to predict the future oil production of the target oil wells. The dynamic data includes: the monthly water injection volume of injection wells and the monthly oil production volume of oil-producing wells in the target area well network; The spatiotemporal neural network model includes: a data preprocessing module, a spatial relationship acquisition module, a time series prediction module, and an output module. The data preprocessing module performs time-delay processing on the filtered historical dynamic data of the injection wells using a one-dimensional convolutional neural network and outputs the resulting labeled data and historical dynamic data to the spatial relationship acquisition module. The spatial relationship acquisition module extracts the spatial features between injection and production wells in the target area well network and outputs them to the time series prediction module. The time series prediction module performs fitting processing on the spatial features using two layers of TCN. The output module outputs the predicted oil production of multiple target oil wells based on the fitting results using multiple single-layer perceptrons. The aforementioned oil well production forecast specifically includes: Step 1) Collect the geometric information of the target well network, construct a graph neural network structure for the injection and production well network based on the topological characteristics of the injection and production wells, and use graph representation to learn and construct the graph data, i.e., the adjacency matrix, which specifically includes: 1.1) The relationship between water injection wells and oil production wells is represented graphically; 1.2) Represent the graph data as an adjacency matrix; 1.3) Save the adjacency matrix to the data preprocessing module; Step 2) Collect the monthly water injection volume of the injection wells and the monthly oil production of the oil production wells in the target area well network; Step 3) Calculate the optimal estimates of monthly water injection volume and monthly oil production volume based on the historical dynamic data of monthly injection wells and production wells using Kalman filtering. This includes: 3.1) Calculate the optimal estimate of the monthly water injection volume at time k. ,in: This represents the actual water injection volume at time k. The observation matrix represents the degree to which the actual system state depends on the calculated data; 3.2) Calculate the Kalman gain Error covariance matrix of monthly water injection volume after Kalman filtering ,in: The observation matrix represents the degree to which the actual system state depends on the calculated data. For noise covariance, Let K be the error covariance moment corresponding to the predicted monthly water injection volume. The Kalman gain is used to iterate through the historical dynamic data of the well network; the filtered data is divided into training and test sets. Step 4) Establish a spatiotemporal neural network model including a data preprocessing module, a spatial relationship acquisition module, a time series prediction module, and an output module. The model uses historical dynamic data of the target area well network processed in Step 3) and the graph data from Step 1) as input, specifically including: Step 4.1) Establish a one-layer one-dimensional convolutional neural network to perform time-lag processing on the monthly water injection volume of the injection well, where the convolutional kernel size is set to n, which means extracting the time-series features between the water injection volumes of n months; Step 4.2) Filter the monthly oil production data of the oil wells, remove the monthly oil production data of the oil wells for the first n-1 months of the data list, and combine it with the dataset required by the water injection data component prediction unit module after convolution. Step 4.3) Perform feature data and label data division on the data group obtained in Step 4.2). Use the data from k time steps as feature data and integrate the oil production data of the target oil well from k+1 time steps as labels. Step 4.4) Use the data obtained in Step 4.3), the graph data, as input to the spatial relationship acquisition module, i.e., a single-layer GCN network layer, and use the ReLU function as the activation function; Step 4.5) The output-input timing prediction module obtained from step 4.4) uses the ReLU function as the activation function; Step 4.6) The output module, which is composed of multiple single-layer perceptrons, obtained from the output of Step 4.5) yields the predicted monthly oil production of the target area oil wells. Step 4.7) Optimize the spatiotemporal neural network model by comparing the monthly oil production of the target wells in the training set with the monthly oil production predicted by the model using the mean square error (MSE). Specifically, the Adam optimization algorithm is used to weight the four loss functions. The weight of the loss function for each well is the proportion of the well's production to the total production of the four wells at k time steps in this training process. ,in: Let MSE be the oil production of the i-th well at time step t. The total loss function is MSE = , These are the mean square errors between the predicted output and the actual value of each of the four wells; Step 5) Predict the monthly oil production of the target oil well using the trained spatiotemporal neural network model, specifically including: Step 5.1) Using the dynamic data of the target well network from the latest k time steps, input the model to predict the monthly oil production of the four oil wells in the future k+1 time step. Step 5.2) Based on the predicted oil production results, adjust the water injection plan for step k+1 and execute the water injection according to this plan to obtain the actual oil production of the target area oil wells in step k+1. Then, incorporate the obtained actual value of step k+1 into the historical data. Step 5.3) Re-input the updated data from the latest k time steps into the trained model to obtain the predicted oil production of the target area wells in the (k+1)th time step of the new period, and repeat the above steps. Step 5.4) Every 50 time steps, select the latest 500 steps of data and re-execute the process after step 35 to achieve continuous updating of the entire prediction model.

2. The oil well production prediction method based on a Kalman filter and a spatiotemporal neural network according to claim 1, characterized in that, The aforementioned time-lag processing refers to: constructing a one-dimensional convolutional network to perform time-lag processing on the monthly water injection volume of injection wells. The convolution kernel size is set to n, representing the extraction and compression of temporal features between the water injection volumes of n months. The convolution direction is from front to back along the time dimension. After convolution, the monthly water injection volume data will be reduced by n-1 time steps in the time dimension. Subsequently, the time-lag processed data is combined with the monthly oil production data of selected oil wells to reconstruct the feature dataset. The first n-1... The next step involves a rejection process. The purpose of this step is to remove the oil production data from the previous n-1 time steps, since the oil production at this moment is the cumulative combined effect of water injection from the injection wells over the previous n months, which is a time lag. The water injection and oil production data processed above are then reintegrated into a single original dataset. At the same time, the prediction dataset is reconstructed, using the data from k time steps as the feature dataset and integrating the oil production data from the target wells at k+1 time steps as the label dataset. Finally, the time series dataset required for prediction is obtained.

3. The oil well production prediction method based on a Kalman filter and a spatiotemporal neural network according to claim 1, characterized in that, The aforementioned extraction of the spatial characteristics between injection and production wells in the target area well network refers to obtaining the spatial relationship between water injection wells and oil production wells in the target area through a single GCN network layer, specifically... ,in: Let A be the feature input for the t-th time node, A be the adjacency matrix representing the graph data of the spatial relationship between well points in the corresponding target area, and D be the degree matrix, whose diagonal elements are obtained by adding the elements of each row of A. This is the corresponding identity matrix.

4. The oil well production prediction method based on a Kalman filter and a spatiotemporal neural network according to claim 1, characterized in that, The fitting process described above is achieved through a two-layer TCN.

5. The oil well production prediction method based on a Kalman filter and a spatiotemporal neural network according to claim 1, characterized in that, The aforementioned graph data refers to: by collecting the geometric information of the target well network, constructing a spatial relationship graph between well points in the injection and production well network based on the topological characteristics of the injection and production wells, and representing it using an adjacency matrix A.

Images