A method for oilfield intrinsic production prediction and risk quantification analysis based on multi-source risk field

Abstract

This invention discloses a method for predicting the intrinsic production of oilfields and quantifying risks based on a multi-source risk field. The method includes: acquiring historical production and multi-source risk data of the oilfield; constructing a unified two-dimensional spatial grid to generate a sequence of historical production and risk heatmaps; performing spatiotemporal alignment and feature fusion on the risk heatmap sequence to form a risk condition field; extracting features and temporally aggregating the production heatmap sequence and the risk field sequence to obtain historical states and risk representations, and fusing them to generate an initial state for prediction; recursively extrapolating future states under risk-driven conditions to decode and generate a future production heatmap sequence; predicting total production and intrinsic production under both true risk and zero risk conditions, and obtaining the risk impact response; and performing attribution analysis using a historical risk sample library to obtain the contribution and proportion of various risks. This invention enables oilfield production prediction and risk impact quantification analysis within a unified spatiotemporal framework, improving the interpretability and reliability of production prediction.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of oilfield development dynamic analysis and artificial intelligence, specifically involving a method for predicting the essential production of oilfields and analyzing risks based on multi-source risk field construction, temporal prediction of state information, and risk quantification attribution. Background Technology

[0002] During oilfield development, oilfield production is constantly changing. With new wells coming online, old wells naturally declining in production, adjustments to intervention well operations, and changes in injection-production systems, the production distribution at different well points and at different times within the oilfield area continuously shifts. Simultaneously, equipment malfunctions, fluctuations in operating conditions, implementation of interventions, changes in water injection, and other production disturbances further influence the spatiotemporal changes in oilfield production. The combined effect of these factors means that future oilfield production includes both fundamental changes determined by formation and development conditions, as well as fluctuations caused by various risks and disturbances, thus posing significant challenges to oilfield production forecasting and risk analysis.

[0003] Most existing oilfield production forecasting methods directly fit or predict future production using historical production data, focusing primarily on improving the accuracy of total production forecasts while neglecting the analysis of the internal components of production changes. Especially during oilfield development, new wells, old wells, and intervention wells coexist, with significant differences in production status between wells. Production evolution within a region exhibits clear spatial distribution and temporal dynamics. Simply treating well point production as independent time series fails to effectively reflect the spatial relationships and temporal patterns within the oilfield region. Furthermore, oilfield development involves a large amount of multi-source risk information, including discrete event risks, continuously monitored risks, and regional statistical risks. These risk information sources differ, are recorded in different ways, have different time frequencies, and different spatial ranges. Existing methods typically use them only as auxiliary information for empirical judgment or employ simple labeling for correlation analysis, lacking a unified spatiotemporal risk expression method that can directly participate in forecast calculations. Therefore, it is difficult to truly depict the dynamic impact of risks on production changes.

[0004] Furthermore, existing methods typically only provide total production forecasts after the combined effects of actual development conditions and risk disturbances, failing to further distinguish between the intrinsic production of an oilfield region under risk-free reference conditions and the additional impacts of risk disturbances. This leads to unclear sources of production fluctuations and difficulty in quantifying risk contributions. Even when some methods can identify anomalies or risk phenomena, they mostly remain at the level of qualitative analysis or ex-post statistics, lacking a quantitative analysis link that unifies total production forecasts, intrinsic production forecasts, and historical similar risk samples. Consequently, it is difficult to generate interpretable decomposition results regarding the degree of impact of different risk types.

[0005] Therefore, how to construct a method that can uniformly express the historical production heat map sequence and the multi-source risk heat map sequence, predict the future total production and the essential production under both real risk and zero risk conditions, and further quantify the impact of risk on production and complete the risk type contribution analysis, has become an urgent technical problem to be solved in the field of dynamic prediction and risk analysis of oilfield development, given the characteristics of the coexistence of historical production distribution and multi-source risk information, dynamic changes and spatial differences during the oilfield development process. Summary of the Invention

[0006] To overcome the above-mentioned shortcomings, this invention proposes a method for predicting the intrinsic production of oilfields and quantifying risks based on multi-source risk fields. The specific steps of this invention are as follows:

[0007] S1: Obtain production data, well location coordinate data, and multi-source risk data for the target oilfield area during historical periods, and construct historical production heat map sequences and multi-source risk heat map sequences on a unified two-dimensional spatial grid;

[0008] S2, perform time alignment, spatial alignment and feature fusion on the multi-source risk heat map sequence to generate a risk condition field sequence corresponding to the target prediction period;

[0009] S3, extracts and fuses features from the historical production heatmap sequence and the risk condition field sequence respectively, to generate initial state information for future prediction;

[0010] S4, taking the initial state information as the starting point for future prediction, and combining the current input information and the risk condition field at the corresponding moment, the future state sequence within the target prediction period is generated recursively through the state prediction model.

[0011] S5, generate a future output heat map sequence based on the future state sequence, calculate the future total output prediction result when the real risk condition field sequence is input, and calculate the essential output prediction result when the risk channel value in the risk condition field sequence is set to zero.

[0012] S6. Calculate the risk impact on output response based on the difference between the future total output forecast and the intrinsic output forecast, and conduct similarity attribution analysis based on the historical risk sample library to obtain the contribution amount and contribution ratio of each risk type.

[0013] S7 outputs the future total output forecast, the intrinsic output forecast, and the risk impact output analysis results.

[0014] The technical features and improvements of this invention are as follows:

[0015] For step S1, the construction process of the unified two-dimensional spatial grid, the historical production heat map sequence, and the multi-source risk heat map sequence specifically includes:

[0016] (1) Construction of a unified two-dimensional spatial grid: Based on the planar coordinate range of each well location within the target oilfield area, determine the lateral coordinate range of the unified two-dimensional spatial grid. and vertical coordinate range and according to the preset grid resolution The target oilfield area is divided into grids to obtain a unified two-dimensional spatial grid. ; wherein, the first in the unified two-dimensional spatial grid The spatial location corresponding to each grid cell is denoted as .

[0017] (2) Construction of historical production heatmap sequence: For each historical moment The production value corresponding to each well point or production unit. According to the well location plane coordinates Mapped to a unified two-dimensional spatial grid The corresponding grid cells in the map are used; for grid cells not directly covered by well points, interpolation is used for numerical completion to obtain the two-dimensional production distribution map corresponding to that historical moment. A historical output heatmap sequence is formed by arranging the two-dimensional output distribution maps corresponding to all historical moments in chronological order. The interpolation result satisfies:

[0018] ;

[0019] In equation (1), Representing historical moments In grid location Estimated output value at the location, Indicates the first Each well point or production unit at a historical moment The output value, Indicates grid position With the The distance between well points This is the distance attenuation coefficient.

[0020] (3) Construction of multi-source risk heat map sequence: For each risk record in the multi-source risk data, extract the risk type. The time of occurrence or the period of effect of the risk Location or center of action of the risk Information on the scope and intensity of the risk; then, based on the location or center of the risk's impact. The diffusion center is defined as the spatial extent corresponding to the risk impact range, within a unified two-dimensional spatial grid. The values ​​of each grid cell within the diffusion area are assigned to generate the risk event at time [time]. The corresponding two-dimensional risk distribution map; for cases where multiple risk events exist at the same time and under the same risk type, the two-dimensional risk distribution maps corresponding to the multiple risk events are superimposed on a unified two-dimensional spatial grid to obtain the risk heatmap corresponding to that risk type at that time. Arrange the risk heatmaps corresponding to each time point and risk type in chronological order to form a multi-source risk heatmap sequence. Among them, a single risk event is located in a grid position. The influence value at point satisfies:

[0021] ;

[0022] In equation (2), Indicates the first The diffusion scale corresponding to the scope of impact of a risk class.

[0023] For step S2, the time alignment, spatial alignment, and feature fusion process of the multi-source risk heatmap sequence specifically includes:

[0024] (1) Time alignment: Using the time scale of the target prediction period as a unified time reference, the risk heat map sequences corresponding to each risk type are resampled; for cases where the original risk heat map recording time is inconsistent with the target prediction time, linear time interpolation is used to map the risk heat map between adjacent known times to the target prediction time, thereby obtaining a risk heat map sequence that corresponds one-to-one with the target prediction period. Among them, the first Risk type at target time The corresponding linear time interpolation result satisfies:

[0025] ;

[0026] In equation (3), and They represent the first Class risk at adjacent known times and The corresponding risk heatmap, Indicates the linear interpolation at the target time. The corresponding risk heatmap.

[0027] (2) Spatial Alignment: Using the unified two-dimensional spatial grid established in step S1 as a unified spatial reference, the risk heatmaps of each risk type after time alignment are resampled; for cases where there are differences in spatial resolution or coverage between the original risk heatmap and the unified two-dimensional spatial grid, bilinear spatial interpolation is used to spatially map the risk heatmap to obtain the unified two-dimensional spatial grid. The risk heatmap shown above Among them, the target grid location The risk value at a given location is calculated by weighting the values ​​of the four adjacent grid points in the original risk heatmap to obtain the interpolation result for that location.

[0028] (3) Feature fusion: fusing features from the same time point The risk heatmaps corresponding to each risk type are stitched together according to the channel dimension to form the multi-channel risk tensor at that moment:

[0029] ;

[0030] In equation (4), Indicates the total number of risk types. This indicates a channel-based splicing operation. Indicates time A multi-channel risk tensor is obtained; then, a convolution fusion operation is performed on the multi-channel risk tensor to obtain the time step. Corresponding risk condition field :

[0031] ;

[0032] In equation (5), The feature fusion operation involves arranging the risk condition fields corresponding to each time point within the entire target prediction period in chronological order to form a risk condition field sequence corresponding to the target prediction period. .

[0033] For step S3, spatial features are extracted from the historical production heatmap sequence at each time step, and then recursively aggregated in chronological order to obtain a historical state representation. Simultaneously, features are extracted from the risk condition field sequence at each time step and then aggregated temporally to obtain a risk characterization. Representing historical states Risk characterization The data is concatenated along the channel dimension, and a predicted initial state is generated through state fusion computation. Its expression is:

[0034] ;

[0035] In equation (6), This indicates a channel-based splicing operation. This indicates a state fusion operation.

[0036] For step S4, the state prediction model used in this invention first processes the current input information. Perform convolutional feature extraction to obtain the current input features. Risk information at the corresponding time point Perform convolutional feature extraction to obtain the current risk features. Then the current status information Current input features and current risk characteristics Concatenate along the channel dimension to form a joint feature tensor. Then, for the joint feature tensor... By sequentially performing multi-layer convolution operations and non-linear activation operations, the state flow information at the current time step is obtained. Finally, the state flow information will be... Current status information Add them together to generate the predicted state information for the next time step. .

[0037] For step S5, firstly, the future state information corresponding to each time point within the target prediction period is processed. Decoding operations are performed separately to obtain the future output heatmap for the corresponding time point. Then, the future output heatmaps corresponding to all predicted times are arranged in chronological order to form a future output heatmap sequence; when the actual risk condition field sequence is input, a future total output heatmap sequence is obtained. The predicted total future output is calculated based on the heatmap sequence of the total future output. After setting the values ​​of each risk channel in the risk condition field sequence to zero, while keeping the historical state information extraction method, state flow calculation method, and decoding method unchanged, a new future output heatmap sequence is generated to obtain the essential output heatmap sequence. The intrinsic yield prediction result is calculated based on the intrinsic yield heatmap sequence. Its expression is:

[0038] ;

[0039] In equation (7), This represents the decoding operation from status information to the production heatmap. This represents the aggregation operation that calculates the total output result from the future output heatmap sequence.

[0040] For step S6, the process of calculating the risk impact on production response and performing risk sample similarity attribution analysis specifically includes:

[0041] (1) Calculation of output response to risk impact: Heat map of future total output at each forecast time and spatial grid location within the target forecast period. Heat map of intrinsic yield By performing time-by-time and grid-by-grid difference calculations, a heatmap of the risk impact on production response corresponding to each prediction time is obtained. And the forecast results for future total output. Compared with the intrinsic yield forecast results The difference is calculated to obtain the overall risk impact on output during the target forecast period. Its expression is:

[0042] ;

[0043] In equation (8), Indicates the first Heatmap of risk impact on output response at each predicted time point This indicates the overall risk impact on output during the target forecast period.

[0044] (2) Risk Sample Similarity Attribution Analysis: The historical risk sample database stores historical risk condition fields, corresponding future total output results, intrinsic output results, and risk impact output response results. First, the spatial distribution characteristics, temporal evolution characteristics, and risk intensity characteristics of the current risk condition field are extracted, and similarity calculations are performed with the corresponding characteristics of each historical sample in the historical risk sample database to obtain the comprehensive similarity between the current risk condition field and each historical sample.

[0045] ;

[0046] In equation (9), Indicates the current risk conditions field and the first The overall similarity between historical samples , and Let represent the weights corresponding to spatial distribution similarity, temporal evolution similarity, and risk intensity similarity, respectively, and satisfy . .

[0047] Then, historical samples with a comprehensive similarity greater than a preset threshold are selected as historical similar risk events. Based on the risk impact response corresponding to the historical similar risk events, the current risk impact on output response is decomposed into risk types to obtain the contribution of each risk type. and contribution percentage Its expression is:

[0048] ;

[0049] In equation (10), This represents the set of historical similar risk samples with a comprehensive similarity greater than a preset threshold. Indicates the first The similarity weights corresponding to each historical sample Indicates the first In the historical sample, the first Risk impact response corresponding to risk class, Indicates the total number of risk types. Indicates the first The contribution percentage of risk categories.

[0050] The oilfield intrinsic production prediction and risk quantification analysis method based on multi-source risk fields of the present invention has the following advantages:

[0051] (1) A unified spatiotemporal representation of historical production information and multi-source risk information has been achieved. By constructing historical production heat map sequences, multi-source risk heat map sequences, and risk condition field sequences, the production distribution and risk distribution within the oilfield area can be expressed at a unified spatial and temporal scale, providing a consistent data foundation for subsequent prediction and analysis.

[0052] (2) It can obtain both future total output and intrinsic output under the same prediction mechanism. Under both real risk and zero risk conditions, the prediction results of future total output and intrinsic output are obtained based on the same state generation process, and the difference between the two is used to quantitatively characterize the impact of risk on output, thus avoiding the problem of incomparable results caused by different models.

[0053] (3) It has the ability to explain risk attribution based on historical similar risk samples. By constructing a historical risk sample library and performing similarity matching, the impact of risk on output can be further decomposed into the contribution amount and contribution ratio of different risk types, which improves the interpretability and engineering guidance significance of risk analysis results.

[0054] (4) The integrated prediction, quantification and attribution method is applicable to complex development scenarios. This invention unifies the prediction of future total production, the prediction of intrinsic production, the quantification of risk impact and the analysis of risk contribution in the same method system, and improves the stability of future state generation through mean flow constraint training. It is applicable to oilfield development scenarios where new wells, old wells and intervention wells coexist and multiple types of risks are constantly changing. Attached Figure Description

[0055] Figure 1 This is the overall flowchart of the oilfield intrinsic production prediction and risk quantification analysis method based on multi-source risk field in this invention.

[0056] Figure 2This is a schematic diagram of the multi-source risk condition field generation structure in this invention.

[0057] Figure 3 This is a schematic diagram of the dual-scenario output prediction structure in this invention.

[0058] Figure 4 This is a schematic diagram of the risk impact attribution analysis process in this invention. Detailed Implementation

[0059] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0060] This invention proposes a method for predicting the intrinsic production of oilfields and quantifying risks based on a multi-source risk field, such as... Figure 1 As shown, its implementation process deconstructs the complex reservoir dynamic analysis task into three core dimensions: "risk condition field generation, dual-scenario state prediction, and historical similarity risk attribution." The following is a detailed description of the specific implementation of each step of this invention, aiming to fully demonstrate the practical application path of this technical solution in engineering.

[0061] S1, Data Acquisition and Spatial Grid Construction. For the target oilfield area, historical production data is first acquired, including daily oil production data of each production well at different times, well location plane coordinates, and corresponding production time information. Assume there are a total of [number missing] wells within the oilfield area. Each production well is at a specific time. The output is recorded as The well location coordinates are as follows . will continue Historical production data from the past 24 days is used as model input. The spatial extent of the oilfield area is determined based on the coordinates of all well locations. and A unified two-dimensional spatial grid is established within this region. In this embodiment, the spatial grid resolution is set to... The oilfield area is divided into 16,384 spatial units, with each grid unit corresponding to a spatial location. Then, each well is analyzed at a specific time... The production data is mapped to the corresponding spatial grid cells according to the well location coordinates, and a two-dimensional production distribution map at each time moment is generated by spatial interpolation method, thus forming a historical production heat map sequence.

[0062] Simultaneously, multi-source risk data generated during oilfield production are acquired, and risk information is mapped to a unified spatial grid based on the time and spatial location of risk occurrence, thereby generating a risk spatial distribution map at the corresponding time, which is then arranged in chronological order to form a multi-source risk heat map sequence.

[0063] S2, Generation of Multi-Source Risk Conditional Field: First, the multi-source risk heatmap sequence is time-aligned. In this embodiment, a daily scale is used as the unified time scale, and the risk heatmaps are resampled over time so that each prediction moment corresponds to a risk spatial distribution map. After completing the spatiotemporal alignment, the risk heatmaps corresponding to different risk types at the same moment are stitched together according to the channel dimension to construct a multi-channel risk tensor. Let the number of risk types be... Then the dimension of the risk tensor at each moment is The multi-channel risk tensor is then input into a risk feature fusion network for feature extraction. The risk feature fusion network consists of three convolutional layers, with each convolutional kernel having a size of [missing information]. The output channel count is 64, and the activation function is ReLU. Through convolutional fusion operations, the spatial correlation between different risk types can be encoded into a unified risk feature representation, forming a risk conditional field.

[0064] S3, State Representation Construction: In this embodiment, spatial features are first extracted from the historical production heatmap sequence time-by-time. The production heatmap corresponding to each historical time point is then constructed. Input convolutional feature extraction module. This module consists of multiple convolutional operations. The first convolutional layer has a kernel size of 3×3 and a stride of 1, used to extract the output variation relationship within the local spatial neighborhood. Subsequently, a second convolutional layer further extracts higher-level spatial distribution features. After the convolutional operations, the output feature map corresponding to that time step is obtained. Temporal aggregation is implemented using a recursive update method. Let the output feature extracted at the current time step be... The historical state at the previous moment was During the calculation process, both are input into the time-series aggregation module for updating, thus obtaining a new historical state representation. This is achieved through continuous... By repeating this calculation process at each historical moment, the historical state representation can be obtained. At the same time, the risk conditions at each moment will be considered. The input risk feature extraction module extracts the spatial distribution features of risk through convolution operations. By sliding the convolution kernel across a two-dimensional spatial grid, the risk values ​​of neighboring grids are weighted and calculated to obtain a risk feature representation. Subsequently, the risk features from consecutive time points are aggregated in chronological order to obtain a risk characterization. .

[0065] S4, risk-driven state recursive prediction, such as Figure 3 As shown, in this embodiment, the input to the state prediction model includes the current state information. Current input information and the corresponding risk conditions at that time Both the input feature extraction branch and the risk feature extraction branch are implemented using convolutional structures, with the convolutional kernel size set to... The convolution stride is 1, and the output feature map spatial size remains consistent with the input grid. The state prediction network consists of multiple convolutional layers and nonlinear activation layers. The first convolutional layer... The system performs feature combination in the local spatial neighborhood, the second convolutional layer further extracts the coupling relationship between multi-channel features, and the last convolutional layer outputs the state flow information at the current time. .

[0066] S5, State Decoding and Dual-Scene Output Calculation: In this embodiment, the state feature map obtained in step S4 is... The input decoding network reconstructs the output spatial distribution map from the state features through convolutional mapping. The decoding network employs a convolutional decoding structure, mapping the feature information in the state space to output distribution values ​​on a unified spatial grid through convolutional operations, thus obtaining the future output heatmap for the corresponding time moment. The real risk condition field sequence is input into the state prediction model, and the future state sequence is generated according to the recursive method of step S4. Subsequently, the state at each time moment is decoded and calculated to obtain the future total output heatmap sequence. Simultaneously, at the model input, the values ​​of each risk channel in the risk condition field sequence are forcibly set to zero. While maintaining the historical state information extraction method, state flow calculation method, and decoding method unchanged, the state recursive calculation in step S4 is re-executed to generate a new future state sequence. This state sequence is then decoded and calculated to obtain the essential output heatmap sequence.

[0067] S6, Risk Impact on Production Response Calculation and Historical Similarity Risk Attribution Analysis: In this embodiment, at the prediction time... At a certain spatial grid location The value of the future total output heatmap is [value]. The intrinsic yield heatmap value is taken as The risk impact response value corresponding to that location is... After completing the calculations for all grid cells in the same manner, the risk impact on production response heatmap at that moment can be obtained. Subsequently, the future total production forecast results can be generated. Compared with the intrinsic yield forecast results The difference is calculated to obtain the overall risk impact on output during the target forecast period. .

[0068] After completing the risk impact response calculation, a similarity attribution analysis of historical risk samples is further performed. First, spatial distribution characteristics are extracted from the current risk condition field. Temporal evolution characteristics and risk intensity characteristics The similarity is then calculated between the current risk condition field and the corresponding features of each sample in the historical sample database. If the comprehensive similarity between the current risk condition field and several historical samples in the sample database is... When it is greater than the preset threshold If so, these samples are selected as historically similar risk events. Then, a similarity weight is calculated based on the similarity of each sample, and a weighted calculation is performed based on the risk impact response corresponding to different risk types in the historical samples, thus obtaining the contribution of each risk type. and its contribution percentage This enables attribution analysis of the impact of current risks on production.

[0069] In summary, this invention proposes a method for predicting the intrinsic production of oilfields and quantifying risks based on a multi-source risk field. By constructing a unified spatial grid and integrating historical production information with multi-source risk information, it achieves spatiotemporal recursive prediction of oilfield production status under risk-driven conditions, and can simultaneously obtain the predicted total future production and the intrinsic production under risk-free conditions. Based on this, by calculating the difference between future total production and intrinsic production, and combining it with historical risk sample similarity attribution analysis, it achieves a quantitative decomposition of the impact degree of different risk types in the oilfield production process. This invention can characterize the spatial diffusion characteristics and temporal delay characteristics of risk impact within a unified spatiotemporal framework, improving the interpretability and reliability of oilfield production prediction results, and providing a scientific basis for oilfield production regulation and risk management.

[0070] Although the present invention has been described in detail through the preferred embodiments above, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above description. Therefore, the scope of protection of the present invention should be defined by the appended claims.

Claims

1. A method for predicting the intrinsic production of oilfields and quantifying risks based on a multi-source risk field, its characteristics and... The specific steps are as follows: S1: Obtain production data, well location coordinate data, and multi-source risk data for the target oilfield area during historical periods, and construct historical production heat map sequences and multi-source risk heat map sequences on a unified two-dimensional spatial grid; S2, perform time alignment, spatial alignment and feature fusion on the multi-source risk heat map sequence to generate a risk condition field sequence corresponding to the target prediction period; S3, extracts and fuses features from the historical production heatmap sequence and the risk condition field sequence respectively, to generate initial state information for future prediction; S4, taking the initial state information as the starting point for future prediction, combines the current input information and the risk condition field at the corresponding moment, and recursively generates the future state sequence within the target prediction period through the state prediction model; S5, Generate a future output heatmap sequence based on the future state sequence, calculate the future total output prediction result when the real risk condition field sequence is input, and calculate the essential output prediction result when the risk channel value in the risk condition field sequence is set to zero. S6. Calculate the risk impact on output response based on the difference between the future total output forecast and the intrinsic output forecast, and conduct similarity attribution analysis based on the historical risk sample library to obtain the contribution amount and contribution ratio of each risk type. S7 outputs the future total output forecast, the intrinsic output forecast, and the risk impact output analysis results. For step S6, the risk sample similarity attribution analysis process specifically includes: the historical risk sample database stores historical risk condition fields, corresponding future total output results, intrinsic output results, and risk impact output response results; firstly, spatial distribution features, temporal evolution features, and risk intensity features are extracted from the current risk condition field, and similarity calculations are performed between these features and the corresponding features of each historical sample in the historical risk sample database to obtain the comprehensive similarity between the current risk condition field and each historical sample: In equation (1), Indicates the current risk conditions field and the first The overall similarity between historical samples , and Let represent the weights corresponding to spatial distribution similarity, temporal evolution similarity, and risk intensity similarity, respectively, and satisfy . ; Then, historical samples with a comprehensive similarity greater than a preset threshold are selected as historical similar risk events. Based on the risk impact response corresponding to the historical similar risk events, the current risk impact on output response is decomposed into risk types to obtain the contribution of each risk type. and contribution percentage Its expression is: In equation (2), This represents the set of historical similar risk samples with a comprehensive similarity greater than a preset threshold. Indicates the first The similarity weights corresponding to each historical sample Indicates the first In the historical sample, the first Risk impact response corresponding to risk class, Indicates the total number of risk types. Indicates the first The contribution percentage of risk categories.

2. The method for predicting oilfield intrinsic production and quantifying risk based on a multi-source risk field as described in claim 1, characterized in that, For step S1, the construction process of the unified two-dimensional spatial grid, the historical production heat map sequence, and the multi-source risk heat map sequence specifically includes: (1) Construction of a unified two-dimensional spatial grid: Based on the planar coordinate range of each well location within the target oilfield area, determine the lateral coordinate range of the unified two-dimensional spatial grid. and vertical coordinate range and according to the preset grid resolution The target oilfield area is divided into grids to obtain a unified two-dimensional spatial grid. ; wherein, the first in the unified two-dimensional spatial grid The spatial location corresponding to each grid cell is denoted as ; (2) Construction of historical production heatmap sequence: For each historical moment The production value corresponding to each well point or production unit. According to the well location plane coordinates Mapped to a unified two-dimensional spatial grid The corresponding grid cells in the map are used; for grid cells not directly covered by well points, interpolation is used for numerical completion to obtain the two-dimensional production distribution map corresponding to that historical moment. A historical output heatmap sequence is formed by arranging the two-dimensional output distribution maps corresponding to all historical moments in chronological order. The interpolation result satisfies: In equation (3), Representing historical moments In grid location The estimated output value at that location, Indicates the first Each well point or production unit at a historical moment The output value, Indicates grid position With the The distance between each well point This is the distance attenuation coefficient; (3) Construction of multi-source risk heat map sequence: For each risk record in the multi-source risk data, extract the risk type. The time of occurrence or the period of effect of the risk Location or center of action of the risk Information on the scope and intensity of the risk; then, based on the location or center of the risk's impact. The diffusion center is defined as the spatial extent corresponding to the risk impact range, within a unified two-dimensional spatial grid. The values ​​of each grid cell within the diffusion area are assigned to generate the risk event at time [time]. The corresponding two-dimensional risk distribution map; for cases where multiple risk events exist at the same time and under the same risk type, the two-dimensional risk distribution maps corresponding to the multiple risk events are superimposed on a unified two-dimensional spatial grid to obtain the risk heatmap corresponding to that risk type at that time. Arrange the risk heatmaps corresponding to each time point and risk type in chronological order to form a multi-source risk heatmap sequence. Among them, a single risk event is located in a grid position. The influence value at point satisfies: In equation (4), Indicates the first The diffusion scale corresponding to the scope of impact of a risk class.

3. The method for predicting the intrinsic production of oilfields and quantifying risks based on a multi-source risk field, as described in claim 1, is characterized in that... For step S2, the time alignment, spatial alignment, and feature fusion process of the multi-source risk heatmap sequence specifically includes: (1) Time alignment: Using the time scale of the target prediction period as a unified time reference, the risk heat map sequences corresponding to each risk type are resampled; for cases where the original risk heat map recording time is inconsistent with the target prediction time, linear time interpolation is used to map the risk heat map between adjacent known times to the target prediction time, thereby obtaining a risk heat map sequence that corresponds one-to-one with the target prediction period. Among them, the first Risk type at target time The corresponding linear time interpolation result satisfies: In equation (5), and They represent the first Class risk at adjacent known times and The corresponding risk heatmap, Indicates the linear interpolation at the target time. The corresponding risk heatmap; (2) Spatial Alignment: Using the unified two-dimensional spatial grid established in step S1 as a unified spatial reference, the risk heatmaps of each risk type after time alignment are resampled; for cases where there are differences in spatial resolution or coverage between the original risk heatmap and the unified two-dimensional spatial grid, bilinear spatial interpolation is used to spatially map the risk heatmap to obtain the unified two-dimensional spatial grid. The risk heatmap shown above Among them, the target grid location The risk value at a given location is calculated by weighting the values ​​of the four adjacent grid points in the original risk heatmap to obtain the interpolation result at that location. (3) Feature fusion: fusing features from the same time point The risk heatmaps corresponding to each risk type are stitched together according to the channel dimension to form the multi-channel risk tensor at that moment: In equation (6), Indicates the total number of risk types. This indicates a channel-based splicing operation. Indicates time A multi-channel risk tensor is obtained; then, a convolution fusion operation is performed on the multi-channel risk tensor to obtain the time step. Corresponding risk condition field : In equation (7), The feature fusion operation involves arranging the risk condition fields corresponding to each time point within the entire target prediction period in chronological order to form a risk condition field sequence corresponding to the target prediction period. .

4. The method for predicting the intrinsic production of oilfields and quantifying risks based on a multi-source risk field as described in claim 1, characterized in that, For step S3, spatial features are extracted from the historical production heatmap sequence at each time step, and then recursively aggregated in chronological order to obtain a historical state representation. Simultaneously, features are extracted from the risk condition field sequence at each time step and then aggregated temporally to obtain a risk characterization. Representing historical states Risk characterization The data is concatenated along the channel dimension, and a predicted initial state is generated through state fusion computation. Its expression is: In equation (8), This indicates a channel-based splicing operation. This indicates a state fusion operation.

5. The method for predicting oilfield intrinsic production and quantifying risk based on a multi-source risk field as described in claim 1, characterized in that, For step S4, the state prediction model used in this invention first processes the current input information. Perform convolutional feature extraction to obtain the current input features. Risk information at the corresponding time point Perform convolutional feature extraction to obtain the current risk features. Then, the current status information is displayed. Current input features and current risk characteristics Concatenate along the channel dimension to form a joint feature tensor. Then, for the joint feature tensor... By sequentially performing multi-layer convolution operations and non-linear activation operations, the state flow information at the current time step is obtained. Finally, the state flow information will be... Current status information Add them together to generate the predicted state information for the next time step. .

6. The method for predicting oilfield intrinsic production and quantifying risk based on a multi-source risk field as described in claim 1, characterized in that, For step S5, firstly, the future state information corresponding to each time point within the target prediction period is processed. Decoding operations are performed separately to obtain the future output heatmap for the corresponding time point. Then, the future output heatmaps corresponding to all predicted times are arranged in chronological order to form a future output heatmap sequence; when the actual risk condition field sequence is input, a future total output heatmap sequence is obtained. The predicted total future output is calculated based on the heatmap sequence of the total future output. After setting the values ​​of each risk channel in the risk condition field sequence to zero, while keeping the historical state information extraction method, state flow calculation method, and decoding method unchanged, a new future output heatmap sequence is generated to obtain the essential output heatmap sequence. The intrinsic yield prediction result is calculated based on the intrinsic yield heatmap sequence. ; Its expression is: In equation (9), This represents the decoding operation from status information to the production heatmap. This represents the aggregation operation that calculates the total output result from the future output heatmap sequence.

7. The method for predicting oilfield intrinsic production and quantifying risk based on a multi-source risk field as described in claim 1, characterized in that, For step S6, the calculation of the risk impact on output response specifically includes: generating a heatmap of the future total output at each prediction time and at each spatial grid location within the target prediction period. Heat map of intrinsic yield By performing time-by-time and grid-by-grid difference calculations, a heatmap of the risk impact on production response corresponding to each prediction time is obtained. And the forecast results for future total output. Compared with intrinsic yield forecast results The difference is calculated to obtain the overall risk impact on output during the target forecast period. Its expression is: In equation (10), Indicates the first Heatmap of risk impact on output response at each predicted time point This indicates the overall risk impact on output during the target forecast period.

Images