Residual oil prediction method based on machine learning

Abstract

The invention relates to the technical field of petroleum engineering, in particular to a remaining oil prediction method based on machine learning. Comprising the steps of multi-source heterogeneous data acquisition, data processing, multi-modal fusion machine learning model construction, landform influence evaluation, oil product sampling detection, prediction evidence based on oil product data, dynamic feedback data acquisition, model dynamic adjustment and iteration and the like. According to the method, through multi-source heterogeneous data acquisition, a multi-modal fusion machine learning model and multi-dimensional influence evaluation on landforms, precise prediction of remaining oil and comprehensive control of environmental risks are realized, so that optimization of a mining scheme is facilitated, and environmental damage is reduced; the method not only can accurately determine the available amount of remaining oil and improve the recovery efficiency, but also can accurately evaluate the influence of mining on terrain, ecology and hydrology and reduce environmental problems.

Description

Technical Field

[0001] The invention relates to the technical field of petroleum engineering, and in particular to a residual oil prediction method based on machine learning. Background Art

[0002] In the oil extraction industry, remaining oil prediction is one of the core links, and its results are directly related to the formulation of extraction plans, resource utilization efficiency and the impact on the surrounding environment.

[0003] Traditional remaining oil prediction methods are mainly based on simple geological models and statistical analysis methods; in the early days, they were mostly based on static data of the reservoir, such as oil layer thickness, porosity, permeability and other parameters to estimate the remaining oil. This method does not fully consider the complexity of the reservoir, so its prediction accuracy is low, usually around 60-70%; for example, in a complex geological structure and highly heterogeneous reservoir, relying solely on these static data is far from accurately reflecting the true distribution of remaining oil; moreover, oil production is a dynamic process, and the pressure, temperature, fluid properties, etc. inside the reservoir will change during production. It is difficult for traditional methods to incorporate these dynamic changes into the prediction system.

[0004] With technological advances, computer simulation technology is used to predict remaining oil. By establishing a numerical simulation model of the reservoir, factors such as fluid flow and reservoir heterogeneity are taken into account to a certain extent. However, this method has obvious shortcomings. On the one hand, it requires a lot of computing resources and precise input parameters. On the other hand, the accuracy of its model is heavily dependent on idealized assumptions about reservoir geology and the production process. More importantly, for oil fields located in areas with complex terrain and landforms, production activities may cause damage to the surrounding terrain, groundwater and ecosystems. Traditional prediction methods can rarely fully evaluate these impacts, making it difficult to avoid environmental risks while ensuring production volumes.

[0005] In summary, the existing remaining oil prediction technology has many limitations in terms of accuracy, comprehensiveness of considerations, data utilization efficiency and data security management. A new and comprehensive remaining oil prediction method is urgently needed to solve these problems. Summary of the invention

[0006] In order to make up for the shortcomings of the prior art, the present invention provides a remaining oil prediction method based on machine learning to overcome the shortcomings of the prior art. The technical solution adopted by the present invention to solve its technical problems is:

[0007] A method for predicting remaining oil based on machine learning comprises the following steps:

[0008] The steps include:

[0009] S1: Multi-source heterogeneous data acquisition: Multi-source heterogeneous data including oilfield geology, mining history, reservoir characteristics, and surrounding topography are collected; geological data include the three-dimensional structure of the oil layer and microscopic pore structure data; mining history data include mining methods and oil production information; reservoir characteristics data include crude oil and formation water chemical parameters; topography data include high-resolution remote sensing and geographic information system (GIS) multi-layer data;

[0010] S2: Data preprocessing: Clean, normalize and reduce the dimension of all data collected in S1 to reduce impurities, unify the scale and complexity of the data, and make the data meet the requirements of subsequent processing;

[0011] S3: Multimodal feature extraction: Use deep learning methods to extract data features collected in S1; for example, use convolutional neural networks to extract texture and structural features from geological image data, and use recurrent neural networks to extract time-dependent features from mining history spatiotemporal series data;

[0012] S4: Construction of multimodal fusion machine learning model: Build a model that integrates multiple machine learning algorithms, including attention-based neural networks, graph convolutional neural networks, long short-term memory networks, etc., and integrate different modal features in S3 through a specific fusion mechanism for subsequent prediction and evaluation;

[0013] S5: Initial remaining oil prediction: The initial data after preprocessing in S2 and feature extraction in S3 are input into the model in S4, and the predicted value of the remaining oil that can be collected is output;

[0014] S6: Topography and geomorphic impact assessment: Evaluate the impact of mining on topography and geomorphology from multiple dimensions, including terrain stability, ecosystem damage risk, and groundwater hydrological change risk, and analyze the potential impact of mining on the surrounding environment;

[0015] S7: Oil sampling and testing: During the mining process, the collected oil is sampled at irregular intervals to test the physical and chemical properties of the oil;

[0016] S8: Prediction support based on oil product data: Integrate the oil product sampling test data in S6 into the model in S4, and verify and optimize the remaining oil prediction results from the perspective of the correlation between the oil product characteristic change trend and the model prediction results;

[0017] S9: Dynamic feedback data collection: Use IoT sensors to collect multi-dimensional data in real time at the mining site to provide a basis for model iteration;

[0018] S10: Dynamic model adjustment and iteration: Based on the incremental learning algorithm, the model is dynamically adjusted and retrained according to the newly collected real-time data, while optimizing the learning rate and using model compression technology to improve model performance.

[0019] Furthermore, the collection of geological data described in S1 includes:

[0020] Three-dimensional geological exploration data, through high-precision seismic exploration technology, obtains three-dimensional spatial distribution information of underground oil layers with sub-meter resolution, including detailed parameters of oil layer boundaries, fault locations, rock layer thickness and geometry, so as to fully describe the structural characteristics of the oil layer;

[0021] Microscopic pore structure image data uses a high-resolution electron microscope combined with micro-CT scanning technology to obtain detailed information on the pores of reservoir rocks at a microscopic scale, including the size range, average size, shape classification (such as spherical, elliptical, irregular), degree of connectivity (measured by the ratio of the number of connected pores to the total number of pores), and the distribution of mineral components on the pore surface and inside, providing a basis for analyzing the oil flow mechanism.

[0022] Further, the dimensionality reduction process in S2 includes the following:

[0023] The three-dimensional geological exploration data and microscopic pore structure image data are subjected to dimensionality reduction processing by combining principal component analysis (PCA) with local linear embedding (LLE); PCA is first used to remove linear correlation in the data, and then LLE is used to maintain the local geometric structure of the data, thereby reducing the data dimension while retaining key geological and pore characteristic information;

[0024] The historical spatiotemporal series data of mining are subjected to dimensionality reduction processing through wavelet transform, the data are decomposed into wavelet coefficients of different frequencies, and the coefficients corresponding to the frequency components that are important for the prediction of remaining oil are selected, thereby reducing the data dimension and highlighting the key time series characteristics;

[0025] The multi-layer data of the geographic information system (GIS) and the high-resolution remote sensing data of topography and landform are subjected to dimensionality reduction processing through feature selection algorithms, such as the ReliefF algorithm, to evaluate the importance of each geographic feature for remaining oil prediction and topography and landform impact assessment, remove unimportant feature dimensions, and retain key features such as terrain elevation, slope, aspect, and land use type.

[0026] Furthermore, feature extraction in S3 includes feature extraction of historical spatiotemporal series data, using the following method:

[0027] Using recurrent neural network (RNN), the temporal attention mechanism is introduced and combined with the gating mechanism to extract features; by setting the update gate, forget gate and output gate, the transmission and update of information in the time series are controlled, so that the model can better capture the long-term dependencies and dynamic change characteristics of data at different stages in the mining process;

[0028] Divide the data at different time scales and perform multi-scale feature extraction on the spatiotemporal series data. For example, aggregate the data at different time intervals such as days, weeks, and months, then extract features separately, and then fuse these multi-scale features to obtain more comprehensive time series features and improve the model's ability to understand the complex change patterns in the mining historical data.

[0029] Furthermore, the multimodal fusion machine learning model in S4 is constructed to calculate the attention weight of each modal feature through a multi-head attention mechanism; the attention calculation paths of multiple heads are set, each head focuses on a different feature subspace, and then the results of multiple heads are spliced ​​together to further enrich and refine the calculation of attention weights, so that the model can more comprehensively capture the complex interactive relationship between different modal features;

[0030] The contribution of different modal features is adjusted through a dynamic weight update strategy. According to the error feedback and data changes during the model training process, the update step and direction of the attention weight are adjusted in real time to ensure that the model can quickly adapt to new data patterns and changes, thereby improving the accuracy and stability of the fusion.

[0031] Furthermore, during the initial remaining oil prediction in S5 and the topographic impact assessment in S6, the model construction for ecosystem damage risk assessment includes:

[0032] Based on vegetation index (such as NDVI) and land cover type data, a vegetation coverage and vegetation health assessment model is constructed. By analyzing the spatiotemporal changes of vegetation index in high-resolution remote sensing data and combining land use type information (such as farmland, forest, grassland, etc.), the direct damage degree and potential impact of mining activities on vegetation are assessed;

[0033] Using reservoir geochemical multi-parameter data and soil type data, a soil pollution diffusion model is established, taking into account the migration and transformation of harmful substances in crude oil (such as heavy metals and organic pollutants) in different soil types, and predicting the scope and degree of soil pollution that may be caused during the mining process, as well as the impact on the growth of surrounding vegetation and ecosystem functions;

[0034] Indicators such as vegetation coverage, vegetation health, and soil pollution level are input into the comprehensive ecological risk assessment model, and the risk level of damage to the surrounding ecosystem caused by mining activities is determined through multi-criteria decision-making methods such as hierarchical analysis method or fuzzy comprehensive evaluation method.

[0035] Furthermore, the sampling point selection criteria for the oil product sampling test in S7 include:

[0036] According to the geological structure of the oil reservoir and the distribution of production wells, sampling points are set at the production wellheads of different oil layers to ensure that the oil products produced in oil layers of different depths are covered to reflect the differences in crude oil characteristics in different areas within the oil reservoir;

[0037] Sampling points are set up at key nodes of the gathering and transportation pipelines, such as the convergence point of the oil pipelines of different oil wells, before and after the oil-water separation equipment, etc., to detect the changes in the properties of oil products during transportation and preliminary treatment, as well as the characteristics of oil products from different oil wells after mixing;

[0038] Dynamic factors in the production process are taken into account. For example, as the production stage progresses, the sampling point position is adjusted in a timely manner according to changes in reservoir pressure, water injection and gas injection effects, etc., to ensure that the sampled oil can accurately represent the remaining oil characteristics under the current production status.

[0039] Furthermore, the specific method of correlation analysis based on prediction evidence of oil product data described in S8 includes:

[0040] Use correlation analysis methods to identify oil parameters that have significant linear relationships with remaining oil reserves, such as density and water content, and determine their correlation direction and strength;

[0041] For nonlinear correlation, data mining technology and machine learning algorithms are used to explore the complex nonlinear mapping relationship between oil parameters and the amount of remaining oil that can be collected, providing a more accurate basis for model prediction;

[0042] By combining a variety of data analysis methods, a comprehensive analysis of multiple oil parameters is conducted to find out the comprehensive characteristics of oil products that can best explain the changes in residual oil.

[0043] Furthermore, the model compression technology in S10 includes:

[0044] Network pruning technology, during the model retraining process, removes neurons and connections that have little impact on the model prediction results based on the importance evaluation indicators of neurons and connections, thereby reducing the complexity and computational complexity of the model.

[0045] Quantization technology converts the weights and activation values ​​in the model from high-precision data representation to low-precision representation. While ensuring that the loss of model accuracy is within an acceptable range, it greatly reduces the resources required for model storage and calculation, and improves the operating efficiency of the model on edge computing devices.

[0046] The beneficial effects of the present invention are as follows:

[0047] 1. The method for predicting remaining oil based on machine learning described in the present invention improves the accuracy of remaining oil prediction through multi-source heterogeneous data collection and multi-modal fusion machine learning models, provides accurate guidance for the optimization of mining plans, avoids excessive or insufficient mining, more accurately determines the amount of remaining oil that can be recovered, reduces resource waste or increased mining costs due to inaccurate predictions, improves recovery rates, and effectively utilizes oil resources.

[0048] 2. The remaining oil prediction method based on machine learning described in the present invention realizes comprehensive environmental risk control through multi-dimensional impact assessment of topography and landforms. It can plan environmental protection measures before mining to reduce damage to the surrounding ecology, and accurately assess terrain stability, ecosystem damage risks and groundwater hydrological change risks, reduce landslides, soil pollution, vegetation destruction and other problems caused by mining, and achieve coordinated development of oil extraction and environmental protection. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The present invention will be further described below in conjunction with the accompanying drawings.

[0050] Figure 1 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0051] In order to make the technical means, creative features, objectives and effects achieved by the present invention easy to understand, the present invention is further explained below in conjunction with specific implementation methods.

[0052] like Figure 1 As shown, the present invention proposes a remaining oil prediction method based on machine learning, which includes the following:

[0053] Multi-source heterogeneous data acquisition: Multi-source heterogeneous data covering oilfield geology, mining history, reservoir characteristics, and surrounding topography are collected from a variety of channels; among them, geological data include the three-dimensional structure of the oil layer and microscopic pore structure data; mining history data covers mining methods, oil production and other information; reservoir characteristic data includes crude oil and formation water chemical parameters; topography data include high-resolution remote sensing and geographic information system (GIS) multi-layer data.

[0054] Data preprocessing: Clean, normalize and reduce the dimension of all collected data to reduce impurities, unify the scale and complexity of the data, and make the data meet the requirements of subsequent processing; let the original data be x, and its minimum value be x min , the maximum value is x max , the normalized formula is For dimensionality reduction, a method combining principal component analysis (PCA) and local linear embedding (LLE) is used for geological exploration three-dimensional data and microscopic pore structure image data. PCA is first used to remove linear correlation in the data. The covariance matrix of the original data is set as ∑. The principal components are obtained by solving the characteristic equation |λI-∑|=0, where λ is the eigenvalue and I is the unit matrix. Then LLE is used to maintain the local geometric structure of the data. For the mining history spatiotemporal series data, wavelet transform is used for dimensionality reduction. The wavelet function is set as ψ(t) and the scaling function is The discrete wavelet transform formula is m, n are scale and translation parameters, a0, b0 are constants, and x(t) is the historical spatiotemporal series data of mining. For multi-layer data of geographic information system (GIS) and high-resolution remote sensing data of topography and landforms, based on feature selection algorithms such as the ReliefF algorithm, the importance of each geographical feature for remaining oil prediction and topography and landform impact assessment is evaluated, and unimportant feature dimensions are removed, while retaining key features such as terrain elevation, slope, aspect, and land use type.

[0055] Multimodal feature extraction: For different types of data, specific deep learning methods are used to extract features; for example, convolutional neural networks are used to extract texture and structural features from geological image data, and recurrent neural networks are used to extract time-dependent features from mining history spatiotemporal series data. In the convolutional neural network, the input data is x ij , i and j represent the location index of the data in the image, and the convolution kernel is k rs , r and s represent the position index of the convolution kernel element, and the convolution kernel size is a×b, then the feature map y pq The calculation is done through convolution operation

[0056] Modal fusion machine learning model construction: Build a model that integrates multiple machine learning algorithms, including neural networks based on attention mechanisms, graph convolutional neural networks, long short-term memory networks, etc., and integrate different modal features through a specific fusion mechanism for subsequent prediction and evaluation; use a multi-head attention mechanism when calculating attention weights, assuming that the query matrix is ​​Q, the key matrix is ​​K, the value matrix is ​​V, and the dimension of the key is d k , the attention weight calculation formula is MultiHead(Q, K, V) = Concat(head1, ..., head h )W O , where h is the number of heads, W O is the output weight matrix, head i Represents the attention calculation result of the ith head; Q, K, and V are obtained by linear transformation of different modal features and are used to measure the correlation between different modal features; in the process of adjusting the contribution of different modal features, a dynamic weight update strategy is introduced to adjust the update step and direction of the attention weight in real time according to the error feedback and data changes in the model training process, to ensure that the model can quickly adapt to new data patterns and changes, and improve the accuracy and stability of fusion.

[0057] Initial remaining oil prediction: Input the initial data after preprocessing and feature extraction into the model, and output the predicted value Q of the remaining oil that can be collected predicted ; Let the input data be X, and the prediction model can be expressed as Q predicted=f(X), where f is a mapping function built based on a variety of machine learning algorithms, used to map input data features to predicted values ​​of the recoverable amount of remaining oil.

[0058] Topography and geomorphology impact assessment: Evaluate the impact of mining on topography and geomorphology from multiple dimensions such as terrain stability, ecosystem damage risk, and groundwater hydrological change risk, and analyze the potential impact of mining on the surrounding environment; in the terrain stability assessment, calculate stress σ and strain ∈ through finite element analysis, and calculate the stability index S based on the specific function S=g(σ,∈), where g is a function determined based on geomechanical principles and experience, and is used to measure the stability of the terrain under the action of mining stress.

[0059] Oil sampling and testing: The collected oil is sampled at key locations and time points in the production process, and scientific sampling methods and advanced testing equipment are used to test multiple physical and chemical properties of the oil. Let a physical or chemical property parameter of the oil be P, and its measurement value is obtained by using the corresponding instrument according to the standard measurement method. For example, the density of the oil is measured by a densitometer, and the chemical composition is measured by instruments such as gas chromatography-mass spectrometry (GC-MS).

[0060] Prediction support based on oil product data: Incorporate oil product sampling test data into the model, and verify and optimize the remaining oil prediction results from the perspective of the correlation between the oil product characteristic change trend and the model prediction results, including trend analysis, correlation analysis and anomaly detection; in the correlation analysis, assume that the oil product parameter is x i , i = 1, 2, ..., n, its mean is The remaining oil that can be collected is y i , whose mean is By calculating the Pearson correlation coefficient Linear correlation is analyzed, and the linear relationship between oil product parameters and the amount of remaining oil that can be recovered is determined based on the size and positive or negative correlation coefficient. For nonlinear correlation, the support vector machine regression (SVR) algorithm is used, with oil product data as input and the amount of remaining oil that can be recovered as output. By training the SVR model, the complex nonlinear mapping relationship between oil product parameters and remaining oil reserves is explored, providing a more accurate basis for model prediction. In addition, the principal component analysis (PCA) and canonical correlation analysis (CCA) methods are combined to conduct a comprehensive analysis of multiple oil product parameters. The principal components of the oil product data are extracted through PCA, and then CCA is used to analyze the correlation between the principal components and the amount of remaining oil that can be recovered, so as to find out the comprehensive characteristics of the oil products that can best explain the changes in the remaining oil.

[0061] Dynamic feedback data collection: Use IoT sensors to collect multi-dimensional data in real time at the mining site to provide a basis for model iteration; suppose the data collected by the downhole pressure sensor is P(t), t is time, P(t) represents the pressure value measured by the downhole pressure sensor at time t, and its value range is determined according to the actual downhole pressure situation; the data collected by the temperature sensor is T(t), which represents the temperature value measured by the temperature sensor at time t, and its value range is determined according to the actual downhole temperature situation, as well as data collected by other types of sensors, such as flow data collected by flow sensors.

[0062] Dynamic adjustment and iteration of the model: Based on the incremental learning algorithm, the model is dynamically adjusted and retrained according to the newly collected real-time data, while optimizing the learning rate and using model compression technology to improve model performance; assuming that the original learning rate is η old , the error change is ΔE, and the learning rate adjustment formula is η new =η old ×h(ΔE), where h is an adjustment function determined based on model training experience and data change characteristics, used to adjust the learning rate according to the error change; in terms of model compression technology, it includes network pruning technology and quantization technology; in network pruning technology, let the activation value of neuron i be a i , with weight w ij (j represents the index of other neurons connected to neuron i), according to a certain threshold θ, (θ is determined according to the model structure and training data, and is used to measure the importance of neurons or connections to the model prediction results), it is determined whether to prune, remove neurons and connections that have little impact on the model prediction results, and reduce the complexity and calculation amount of the model; in the quantization technology, the quantization function is set to (s is the quantization step size, which is determined according to the data type and accuracy requirements of the model parameters). The weights and activation values ​​in the model are converted from high-precision data representations (such as 32-bit floating-point numbers) to low-precision representations (such as 8-bit integers). While ensuring that the loss of model accuracy is within an acceptable range, the resources required for model storage and calculation are greatly reduced, thereby improving the operating efficiency of the model on edge computing devices.

[0063] As a specific implementation of the present invention, the collection of geological data includes:

[0064] For geological exploration three-dimensional data, high-precision seismic exploration technology is used to obtain three-dimensional spatial distribution information of underground oil layers with sub-meter resolution, including detailed parameters such as oil layer boundaries, fault locations, rock layer thickness and geometric shapes, so as to fully describe the structural characteristics of the oil layer; assuming that the coordinate ranges of the oil layer in the x, y, and z directions are [x min , x max ]、[y min ,y max ]、[zmin , z max ], then the oil layer volume This integral is used to calculate the volume of the oil layer in three-dimensional space and provide basic data for related analyses such as reservoir reserve estimation, where dx, dy, and dz represent tiny length increments in the x, y, and z directions.

[0065] For the microscopic pore structure image data, a high-resolution electron microscope combined with micro-CT scanning technology is used to obtain detailed information on the pores of reservoir rocks at the microscopic scale, including the size range, average size, shape classification (such as spherical, elliptical, irregular), degree of connectivity (measured by the ratio of the number of connected pores to the total number of pores), and the distribution of mineral components on the pore surface and inside. Assume that the number of connected pores is n c , the total number of pores is n t , degree of connectivity This ratio is used to describe the degree of interconnection between pores and is of great significance for analyzing the flow characteristics of oil in rocks.

[0066] As a specific implementation of the present invention, the specific method of dimensionality reduction processing in the data preprocessing step is:

[0067] For geological exploration three-dimensional data and microscopic pore structure image data, a method combining principal component analysis (PCA) and local linear embedding (LLE) is used. PCA is first used to remove linear correlation in the data. The original data covariance matrix ∑ is a matrix composed of the covariance between data vectors, and its elements σ ij Represents the covariance of the i-th and j-th data vectors. The principal components are obtained by solving the characteristic equation |λI-∑|=0 (λ is the eigenvalue, I is the unit matrix, and the unit matrix is ​​a square matrix with the main diagonal elements being 1 and the remaining elements being 0). These principal components are the projections of the original data on the new linearly independent dimensions. Then, LLE is used to maintain the local geometric structure of the data, thereby retaining key geological and pore characteristic information while reducing the data dimension. LLE maintains the local geometric structure of the data by finding the local linear relationship of each data point, so that the reduced-dimensional data has a similar structure to the original data locally.

[0068] For the historical spatiotemporal series data of mining, wavelet transform is used for dimensionality reduction. By decomposing the data into wavelet coefficients of different frequencies, the coefficients corresponding to the frequency components that are important for the prediction of remaining oil are selected, thereby reducing the data dimension and highlighting the key time series characteristics. The wavelet function ψ(t) is a function with localization characteristics, which has good localization performance in both the time domain and the frequency domain. The scale function Used to generate wavelet basis, discrete wavelet transform formula In the expression, m and n respectively control the scale and position of the wavelet, a0 and b0 are constants, and x(t) is the historical spatiotemporal sequence data. By changing m and n, wavelet coefficients of different scales and positions can be obtained, thereby achieving data decomposition and dimensionality reduction.

[0069] For geographic information system (GIS) multi-layer data and high-resolution remote sensing data of topography and geomorphology, based on feature selection algorithms, such as the ReliefF algorithm, the importance of each geographical feature for remaining oil prediction and topography and geomorphology impact assessment is evaluated, and unimportant feature dimensions are removed, while retaining key features such as terrain elevation, slope, slope aspect, and land use type; the ReliefF algorithm evaluates the importance of features by calculating the ability of each feature to distinguish sample classification (here it can be understood as different categories of remaining oil prediction and topography and geomorphology impact assessment). Features with strong distinguishing ability are retained, while features with weak distinguishing ability are considered to contribute less to prediction and assessment and are removed.

[0070] As a specific implementation of the present invention, the feature extraction optimization for mining historical spatiotemporal series data includes:

[0071] When using recurrent neural network (RNN) to extract features, in addition to introducing the temporal attention mechanism, a gating mechanism is also combined; by setting the update gate zt and the forget gate r t and output gate o t , controls the transmission and update of information in the time series, so that the model can better capture the long-term dependencies and dynamic change characteristics of data at different stages in the mining process; let the current time be t, and the hidden state at the previous time be h t-1 , the current input is x t , the weight matrices are W z , W r , W h , the bias vectors are b z , b r , b h , the activation function is σ (σ can be a common sigmoid function, used to map the input to the (0, 1) interval), and the nonlinear function is (For example, it can be the hyperbolic tangent function tanh), the calculation formula is z t =σ(W z [h t-1 , x t ]+b z ), r t =σ(W r [h t-1 , x t ]+b r ), Where ⊙ represents element-wise multiplication; update gate z tDetermines the proportion of new input information in the current state update, the forget gate r t Determines how much past information to forget, output gate o t This mechanism can better capture the long-term dependencies and dynamic change characteristics of data at different stages of the mining process.

[0072] Perform multi-scale feature extraction on spatiotemporal series data; divide the data at different time scales, such as aggregating data at different time intervals such as days, weeks, and months, and then extract features separately, and then fuse these multi-scale features to obtain more comprehensive time series features, and improve the model's ability to understand complex change patterns in historical mining data; when aggregating data at different time scales, group the data according to time intervals. For example, daily aggregation is to treat each day's data as a group, and then extract features for each group of data. The features extracted at these different scales contain information at different time granularities, and after fusion, they can more comprehensively describe the change pattern of historical mining data.

[0073] As a specific implementation of the present invention, in the construction of a multimodal fusion machine learning model, the details of the fusion method based on the attention mechanism are as follows:

[0074] When calculating the attention weight of each modal feature, a multi-head attention mechanism is used; by setting the attention calculation path of multiple heads, each head focuses on different feature subspaces, and then splicing the results of multiple heads together, the calculation of attention weights is further enriched and refined, so that the model can more comprehensively capture the complex interactive relationship between different modal features; the attention calculation of each head is performed independently, and the number of heads h is determined according to the complexity of the model and the diversity of data features. In the multimodal fusion scenario, different modal features (such as geological features, mining history features, topographic features, etc.) are linearly transformed to obtain their own Q, K, and V matrices. The dimensions and element values ​​of these matrices are determined by the dimensions of the modal features and the data content; for example, for geological feature modalities, each row vector of its Q matrix represents the query information of geological features for other modal features in a certain dimension, and the column vector of the K matrix is ​​the key value used to match the query information. The attention weight is obtained by calculation, which reflects the correlation between different modalities in this subspace.

[0075] In the process of adjusting the contribution of different modal features, a dynamic weight update strategy is introduced. According to the error feedback and data changes during model training, the update step and direction of the attention weight are adjusted in real time; the error feedback information is set to ΔE t (t represents the training round), the data change is represented by the difference ΔD between the newly collected data and the historical data; the step size α and direction are updated According to ΔE tand ΔD are determined by the following formula: α = g(ΔE t , ΔD), (g and h are functions designed based on model characteristics and training experience) to ensure that the model can quickly adapt to new data patterns and changes, and improve the accuracy and stability of fusion. This dynamic update mechanism enables the model to more accurately adjust the importance of different modal features in the fusion process when facing complex and changeable data in the process of oil reservoir exploitation (such as data changes caused by geological structure changes, mining process adjustments, etc.).

[0076] As a specific implementation of the present invention, in the initial remaining oil prediction and topographic impact assessment, the model construction of ecosystem damage risk assessment includes:

[0077] Based on vegetation index (such as NDVI) and land cover type data, a vegetation coverage and vegetation health assessment model is constructed; by analyzing the spatiotemporal changes of vegetation index in high-resolution remote sensing data and combining land use type information (such as farmland, forest, grassland, etc.), the direct damage degree and potential impact of mining activities on vegetation are assessed; NDVI is set i is the vegetation index of the ith pixel, and its value is calculated through remote sensing data. Its range is generally between [-1, 1]. Positive values ​​indicate vegetation coverage, and the larger the value, the denser and healthier the vegetation. (V is the vegetation pixel set, N V The vegetation coverage is obtained by calculating the ratio of the sum of NDVI of all vegetation pixels in the area to the total number of vegetation pixels. This indicator directly reflects the coverage of vegetation in the area. In addition to coverage, the time series changes of vegetation index can also be combined to evaluate the health of vegetation. For example, a continuous decline in vegetation index may indicate that the vegetation is stressed or damaged.

[0078] A soil pollution diffusion model is established using reservoir geochemical multi-parameter data and soil type data. The migration and transformation of harmful substances in crude oil (such as heavy metals and organic pollutants) in different soil types are considered to predict the scope and degree of soil pollution that may be caused during the mining process, as well as the impact on the growth of surrounding vegetation and ecosystem functions. The pollutant concentration C (x, y, t) in the soil satisfies the diffusion equation: (D diffusion coefficient, its value depends on the physical properties of the soil (such as porosity, permeability, etc.) and the nature of the pollutant, v x 、v yis the velocity component, and its direction and magnitude are related to factors such as soil slope, groundwater flow direction and velocity. R is the reaction term, which indicates the rate of chemical reaction generation or consumption of pollutants in the soil, and is related to the chemical properties of the soil (such as pH, redox potential, etc.) and the type of pollutants. The soil type and initial pollutant concentration are determined by sampling and analyzing the soil around the oil reservoir, and D and v are determined by combining information such as topography and groundwater flow. x 、v y The value of is determined, and R is determined according to the chemical properties of the pollutants in the crude oil. The diffusion equation is solved to obtain the pollutant concentration C(x, y, t) at different times t and spatial positions (x, y).

[0079] The vegetation coverage, vegetation health, soil pollution level and other indicators are input into the comprehensive ecological risk assessment model, and the risk level of mining activities on the surrounding ecosystem is determined through multi-criteria decision-making methods such as hierarchical analysis method or fuzzy comprehensive evaluation method. In the hierarchical analysis method, a hierarchical structure model is first constructed, with the ecosystem damage risk as the target layer and the vegetation coverage, vegetation health, soil pollution level and other indicators as the criterion layer. Then, by constructing a judgment matrix (determining the relative importance of each criterion based on expert experience or historical data), the weight vector of each criterion is calculated, and finally, the indicators under each criterion are quantitatively scored and weighted summed to obtain the comprehensive evaluation value of the ecosystem damage risk. In the fuzzy comprehensive evaluation method, the evaluation factor set (such as vegetation coverage, vegetation health, soil pollution level, etc.) and the evaluation level set (such as low risk, medium risk, high risk) are determined, and a fuzzy relationship matrix (representing the fuzzy relationship between the evaluation factors and the evaluation level) is established. The fuzzy evaluation result of the ecosystem damage risk is obtained through fuzzy synthesis operation, and then the risk level is determined according to the maximum membership principle.

[0080] As a specific implementation of the present invention, in oil sampling and testing, the sampling points are selected based on:

[0081] According to the geological structure of the reservoir and the distribution of production wells, sampling points are set at the production wellheads of different oil layers to ensure that the oil products produced by oil layers of different depths are covered to reflect the differences in crude oil characteristics in different areas within the reservoir. Suppose there are n oil layers of different depths in the reservoir, and the depth of the i-th oil layer is h i , then the sampling points should cover different h i Corresponding wellhead; oil layer depth h iThe sampling frequency of each wellhead is determined by geological exploration data and the production rate and dynamic changes of the reservoir. For example, the sampling frequency can be appropriately increased for wellheads of oil layers with fast production rates or large changes in reservoir properties. By sampling at wellheads of oil layers at different depths, information on changes in crude oil properties in the vertical direction of the reservoir can be obtained, such as density, viscosity, chemical composition, etc., which may change with depth. This information is crucial for a comprehensive understanding of the characteristics of crude oil in the reservoir and accurate prediction of remaining oil conditions.

[0082] Sampling points are set up at key nodes of gathering and transportation pipelines, such as the convergence points of oil pipelines from different oil wells, before and after oil-water separation equipment, etc., to detect changes in the properties of oil products during transportation and preliminary treatment, as well as the characteristics of oil products from different oil wells after mixing; sampling at the convergence points of oil pipelines can analyze changes in the physical and chemical properties of oil products from different oil wells after mixing, because crude oil from different oil wells may have different characteristics; sampling before and after oil-water separation equipment can monitor changes in water content of oil products during the separation process, impurity removal, etc.; the sampling frequency is determined according to factors such as the oil flow rate in the pipeline and the frequency of changes in the treatment process. For example, for pipeline nodes with large flow rates or frequent adjustments to the treatment process, the sampling frequency can be increased accordingly to ensure that changes in oil properties are grasped in a timely and accurate manner.

[0083] Dynamic factors in the production process are taken into account. For example, as the production stage progresses, the sampling point position is adjusted in a timely manner according to changes in reservoir pressure, water injection and gas injection effects, etc., to ensure that the sampled oil can accurately represent the remaining oil characteristics under the current production status; when the reservoir pressure drops, dissolved gas may precipitate from the crude oil, affecting the properties of the oil. At this time, sampling points can be added in areas that may be affected (such as pipelines near oil layers with large pressure changes); during the water injection and gas injection process, the injected material may interact with the crude oil and change the composition and properties of the oil. Therefore, the sampling point position and sampling frequency are appropriately adjusted in the pipeline or wellhead within the influence range of water injection and gas injection to ensure that the sampled oil can reflect the remaining oil characteristics under such dynamic changes.

[0084] As a specific implementation of the present invention, in the prediction evidence based on oil product data, the specific method of correlation analysis includes:

[0085] First, the Pearson correlation coefficient is used to analyze the linear correlation between the physical and chemical properties of oil products and the recoverable amount of remaining oil; the oil parameters with significant linear relationship with the remaining oil reserves, such as density and water content, are identified, and their correlation direction and strength are determined; the oil parameter vector is assumed to be X = (x1, x2, ..., x n ), the amount of remaining oil that can be collected is y, n is the number of oil parameters; the mean (m is the number of samples) Pearson correlation coefficient The value range of r is between [-1, 1], r>0 indicates positive correlation, that is, when the oil parameters increase, the amount of remaining oil that can be recovered also increases; r<0 indicates negative correlation, such as density and the amount of remaining oil that can be recovered may be negatively correlated, that is, the greater the density, the less the amount of remaining oil that can be recovered may be; r=0 indicates no linear correlation; by calculating the values ​​between different oil parameters and the amount of remaining oil that can be recovered, the key parameters that have a linear impact on the prediction of remaining oil can be screened out.

[0086] For nonlinear correlation, support vector machine regression (SVR) algorithm is used; oil product data is used as input, and the remaining oil recoverable amount is used as output. By training the SVR model, the complex nonlinear mapping relationship between oil product parameters and remaining oil reserves is explored to provide a more accurate basis for model prediction; in the SVR model, the input oil product data is x = (x1, x2, ..., x n ), through the kernel function K(x, x i ) (Common kernel functions include linear kernel, polynomial kernel, Gaussian kernel, etc., which are selected according to the characteristics of the data) Map the input data to a high-dimensional feature space and build a linear regression model in the high-dimensional space (ω is the weight vector, The model is trained by minimizing the loss function (such as ∈-insensitive loss function) to find the optimal ω and b, thereby establishing a nonlinear mapping relationship between oil product parameters and the amount of remaining oil that can be collected.

[0087] Combining principal component analysis (PCA) and canonical correlation analysis (CCA) methods, a comprehensive analysis of multiple oil parameters is performed; the principal components of oil data are extracted by PCA, and the oil data matrix is ​​set as X, and its covariance matrix is ​​C. By solving the eigenvalues ​​and eigenvectors of C, the eigenvectors are sorted from large to small according to the corresponding eigenvalues, and the first k eigenvectors are selected to form the projection matrix P, then the principal component Z = XP; then CCA is used to analyze the correlation between the principal component and the amount of remaining oil that can be collected, and the principal component matrix is ​​set as Z, and the vector of the amount of remaining oil that can be collected is y. By calculating the canonical correlation coefficient and canonical variable between Z and y, the comprehensive characteristics of oil products that can best explain the changes in remaining oil are found, and the prediction ability of the model for remaining oil is further optimized.

[0088] As a specific implementation of the present invention, in the dynamic adjustment and iteration of the model, the specific application of the model compression technology includes:

[0089] Using network pruning technology, during the model retraining process, neurons and connections with little impact on the model prediction results are removed according to the importance evaluation indicators of neurons and connections, reducing the complexity and calculation amount of the model; assuming that the activation value of neuron i is a i , with weight w ij(j represents the index of other neurons connected to neuron i), and whether to prune is determined based on a certain threshold θ; the threshold θ can be determined by a variety of methods. One is based on statistical analysis. In the early stage of model training, the activation values ​​and weight distributions of different neurons and connections are observed, and an initial threshold is determined based on the statistical characteristics of the distribution (such as mean, standard deviation, etc.), and then dynamically adjusted according to changes in model performance during the training process. If the model overfits at a certain stage, the threshold can be appropriately lowered and the pruning intensity can be increased; if the model performance decreases too quickly, the threshold can be appropriately increased; another method is to add an auxiliary loss function to integrate the importance evaluation of neurons and connections into the overall training goal of the model, and automatically adjust the threshold based on the optimization result of the loss function; in the pruning process, for neurons and connections whose activation values ​​are close to zero for a long time or whose absolute weights are small, it is considered that they have little impact on the model prediction results and can be deleted, thereby simplifying the model structure and improving the computational efficiency and generalization ability of the model.

[0090] Use quantization technology to convert the weights and activation values ​​in the model from high-precision data representation (such as 32-bit floating point numbers) to low-precision representation (such as 8-bit integers); let the quantization function be (s is the quantization step size), the selection of the quantization step size s is determined according to the data type, numerical range and accuracy requirements of the model parameters; for the weight parameters, a suitable s value can be selected by analyzing their distribution range during the training process, so that the quantized weights can minimize the data storage and calculation amount while ensuring that the loss of model accuracy is within an acceptable range; for the activation values, since they may have different distribution characteristics in different layers and different training stages, an adaptive quantization method can be used to adjust the quantization step size according to the dynamic changes of the activation values ​​of each layer; for example, when the activation value range of some layers is narrow, a smaller quantization step size is selected to ensure quantization accuracy; when the activation value range is wide, the quantization step size is appropriately increased, and some compensation mechanisms (such as quantization-aware training) are used to reduce the impact of quantization errors on model performance, thereby maintaining the prediction accuracy and stability of the model while compressing the model, and improving the operating efficiency of the model on edge computing devices.

[0091] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for predicting remaining oil based on machine learning, characterized in that: The steps include: S1: Multi-source heterogeneous data acquisition: Collect multi-source heterogeneous data covering oilfield geology, mining history, reservoir characteristics, and surrounding topography. Among them, geological data includes the three-dimensional structure of the oil layer and microscopic pore structure data; mining history data covers mining methods and oil production information; reservoir characteristics data includes crude oil and formation water chemical parameters; topography data includes high-resolution remote sensing and geographic information system multi-layer data; S2: Data preprocessing: cleaning, normalizing and dimension reduction of all data collected in S1; S3: Multimodal feature extraction: Use deep learning methods to extract features of the data collected in S1; S4: Multimodal fusion machine learning model construction: Build a model that integrates multiple machine learning algorithms and integrates the different modal features in S3 for subsequent prediction and evaluation; S5: Initial remaining oil prediction: The initial data after preprocessing in S2 and feature extraction in S3 are input into the model in S4, and the predicted value of the remaining oil that can be collected is output; S6: Topography and geomorphic impact assessment: Evaluate the impact of mining on topography and geomorphology from multiple dimensions, including terrain stability, ecosystem damage risk, and groundwater hydrological change risk, and analyze the potential impact of mining on the surrounding environment; S7: Oil sampling and testing: During the mining process, the collected oil is sampled at irregular intervals to test the physical and chemical properties of the oil; S8: Prediction support based on oil product data: Integrate the oil product sampling test data in S6 into the model in S4, and verify and optimize the remaining oil prediction results from the perspective of the correlation between the oil product characteristic change trend and the model prediction results; S9: Dynamic feedback data collection: Use IoT sensors to collect multi-dimensional data in real time at the mining site to provide a basis for model iteration; S10: Dynamic model adjustment and iteration: Based on the incremental learning algorithm, the model is dynamically adjusted and retrained according to the newly collected real-time data, while optimizing the learning rate and using model compression technology to improve model performance.

2. The method for predicting residual oil based on machine learning according to claim 1, characterized in that: The collection of geological data described in S1 includes: Geological exploration three-dimensional data, that is, the three-dimensional spatial distribution information of underground oil layers with sub-meter resolution. Microscopic pore structure image data, that is, detailed information of reservoir rock pores at the microscopic scale, provides a basis for analyzing oil flow mechanisms.

3. The method for predicting residual oil based on machine learning according to claim 2, characterized in that: The dimensionality reduction process in S2 includes the following: The three-dimensional geological exploration data and microscopic pore structure image data are subjected to dimensionality reduction processing by combining principal component analysis with local linear embedding; The mining history spatiotemporal sequence data is subjected to dimensionality reduction processing by wavelet transform; The multi-layer data of the geographic information system and the high-resolution remote sensing data of topography and landform are subjected to dimensionality reduction processing through a feature selection algorithm.

4. The method for predicting residual oil based on machine learning according to claim 1, characterized in that: Feature extraction in S3 includes feature extraction of historical spatiotemporal series data, using the following methods: Use recurrent neural networks, introduce temporal attention mechanisms, and combine gating mechanisms to extract features; The data is divided into different time scales, and multi-scale features are extracted from spatiotemporal series data.

5. The method for predicting residual oil based on machine learning according to claim 1, characterized in that: In S4, a multimodal fusion machine learning model is constructed, and the attention weight of each modality feature is calculated through a multi-head attention mechanism; And the contribution of different modal features is adjusted through a dynamic weight update strategy.

6. The method for predicting residual oil based on machine learning according to claim 1, characterized in that: During the initial remaining oil prediction in S5 and the topographic impact assessment in S6, the model construction of ecosystem damage risk assessment includes: Based on vegetation index and land cover type data, a vegetation coverage and vegetation health assessment model was constructed; Using reservoir geochemical multi-parameter data and soil type data, a soil contamination diffusion model was established; Indicators such as vegetation coverage, vegetation health, and soil pollution level are input into the comprehensive ecological risk assessment model, and the risk level of damage to the surrounding ecosystem caused by mining activities is determined through analysis.

7. The method for predicting residual oil based on machine learning according to claim 1, characterized in that: The basis for selecting the sampling points for oil sampling and testing described in S7 includes: According to the geological structure of the oil reservoir and the distribution of production wells, sampling points are set at the production wellheads of different oil layers and at key nodes of the gathering and transportation pipelines.

8. The method for predicting residual oil based on machine learning according to claim 1, characterized in that: The model compression technology described in S10 includes: network pruning technology and quantization technology, and the quantization technology is used to convert the weights and activation values ​​in the model from a high-precision data representation form to a low-precision representation form.

Images