Coalbed methane production capacity prediction method and system combined with artificial intelligence

Abstract

The present invention provides a coalbed methane production capacity prediction method and system combined with artificial intelligence. First, a multi-source coalbed methane production capacity data set including geological structure data, reservoir physical property data and historical production dynamic data of the target block is obtained. Then, spatiotemporal alignment processing is performed on the multi-source coalbed methane production capacity data set to obtain a standardized coalbed methane production capacity feature set. Thereafter, feature extraction is performed on the standardized coalbed methane production capacity feature set to generate a prediction feature set including dual encoding of geological attribute features and dynamic attribute features. Then, the prediction feature set is iteratively predicted based on a reinforcement learning strategy network to obtain a target coalbed methane production capacity prediction result. Finally, an optimization strategy set is generated according to the error distribution between the target coalbed methane production capacity prediction result and the historical production dynamic data, and the strategy is fed back to the coalbed methane production control system to trigger a parameter adjustment operation, thereby improving the accuracy of the coalbed methane production capacity prediction and realizing optimized control of the production process.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a coalbed methane production capacity prediction method and system combined with artificial intelligence. Background Art

[0002] As an important unconventional natural gas resource, coalbed methane (CBM) production capacity forecasting is crucial for its efficient development and rational utilization. In the field of traditional CBM production capacity forecasting, early methods failed to fully consider the temporal and spatial characteristics of data during data processing. Temporal and spatial inconsistencies in data from different sources made it difficult to effectively correlate and integrate data, which in turn affected subsequent analysis and forecasting.

[0003] In addition, most traditional methods are based on simple statistical analysis or fixed mathematical models, which cannot dynamically adapt to the ever-changing conditions during the coalbed methane production process. These models lack the ability to mine the underlying complex relationships in the data and cannot effectively handle the interaction between geological attribute characteristics and dynamic attribute characteristics, resulting in difficulty in accurately predicting coalbed methane production capacity in practical applications. Moreover, when the predicted results deviate from the actual production situation, traditional methods lack an effective feedback mechanism to adjust production parameters in a timely manner, making it impossible to achieve optimal control of the coalbed methane production process. Therefore, existing coalbed methane production capacity prediction methods have shortcomings in terms of accuracy, adaptability, and optimized control. There is an urgent need for a more advanced and efficient prediction method to meet the needs of actual production. Summary of the Invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a coalbed methane production capacity prediction method combined with artificial intelligence, the method comprising:

[0005] Acquire a multi-source coalbed methane production capacity data set for a target block, wherein the multi-source coalbed methane production capacity data set includes geological structure data, reservoir physical property data, and historical production performance data;

[0006] Performing spatiotemporal alignment processing on the multi-source coalbed methane production capacity data set to generate a standardized coalbed methane production capacity feature set;

[0007] Extracting features from the standardized coalbed methane productivity feature set to generate a prediction feature set, wherein each feature unit in the prediction feature set includes dual coding of geological attribute features and dynamic attribute features;

[0008] Iteratively predict the geological attribute characteristics and dynamic attribute characteristics based on the reinforcement learning strategy network to generate a target coalbed methane production capacity prediction result;

[0009] An optimization strategy set is generated according to the error distribution between the target coalbed methane production capacity prediction result and the historical production dynamic data, and the optimization strategy set is fed back to the coalbed methane production control system to trigger a parameter adjustment operation.

[0010] On the other hand, an embodiment of the present invention also provides a coalbed methane production capacity prediction system combined with artificial intelligence, including a processor and a machine-readable storage medium, the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0011] Based on the above aspects, the embodiment of the present invention obtains a multi-source coalbed methane production capacity data set containing geological structure, reservoir physical properties and historical production dynamic data, and performs spatiotemporal alignment processing to generate a standardized coalbed methane production capacity feature set, which effectively solves the problem of difficulty in integrating spatiotemporal inconsistencies in multi-source data. Feature extraction is performed on the standardized coalbed methane production capacity feature set to generate a prediction feature set containing dual encoding of geological attribute features and dynamic attribute features, which can comprehensively and deeply explore the potential complex relationship between geological attributes and dynamic attribute features in the data. The prediction feature set is iteratively predicted based on the reinforcement learning strategy network, which can dynamically adapt to changes in the coalbed methane production process. Compared with traditional fixed models, it can more accurately generate target coalbed methane production capacity prediction results. According to the error distribution of the target coalbed methane production capacity prediction result and the historical production dynamic data, an optimization strategy set is generated and fed back to the coalbed methane production control system to trigger parameter adjustment operations, thereby achieving improved accuracy of coalbed methane production capacity prediction and optimized control of the production process. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 This is a schematic diagram of the execution flow of the coalbed methane production capacity prediction method combined with artificial intelligence provided by an embodiment of the present invention.

[0013] Figure 2 Schematic diagram of exemplary hardware and software components of a coalbed methane production capacity prediction system combined with artificial intelligence provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0014] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of a coalbed methane production capacity prediction method combined with artificial intelligence provided by an embodiment of the present invention. The coalbed methane production capacity prediction method combined with artificial intelligence is introduced in detail below.

[0015] Step S110: Acquire a multi-source coalbed methane production capacity data set of a target block, wherein the multi-source coalbed methane production capacity data set includes geological structure data, reservoir physical property data, and historical production dynamic data.

[0016] In this embodiment, geological structure data is key information reflecting the geological environment in which the coal seam is located. For example, through seismic exploration technology, the propagation characteristics of seismic waves in underground media are utilized to analyze reflected waves, refracted waves, etc., thereby obtaining information such as the structure and morphology of the stratum. This information can be used to determine parameters such as the stratum dip and fault location; drilling is a method of directly obtaining underground core samples. Through observation and analysis of the core, the rock type, fracture development, etc. of the stratum can be more accurately understood.

[0017] Reservoir physical data relates to the physical characteristics of coal seams, such as porosity and permeability, which directly impact the storage and flow capacity of coalbed methane. Reservoir physical data is typically obtained through professional laboratory analysis of collected core samples. Relevant parameters are determined by measuring indicators such as pore volume and fluid flow capacity.

[0018] Historical production dynamics data records past CBM production conditions, including gas production volume sequences, pressure change sequences, and production cycle phase labels. This data can be collected from various monitoring devices deployed during CBM production, such as gas flow sensors and pressure sensors. These devices record various production data in real time and store them in a database for subsequent analysis.

[0019] By collecting and integrating the above multi-source data, a multi-source CBM production capacity data set is formed, which includes geological structure data, reservoir physical property data and historical production dynamic data.

[0020] Step S120: performing spatiotemporal alignment processing on the multi-source coalbed methane productivity data set to generate a standardized coalbed methane productivity feature set.

[0021] In this example, because the multi-source CBM production capacity data set comes from diverse sources, with varying acquisition times and spatial ranges, and differing data formats and precision, spatiotemporal alignment is required to ensure data consistency and comparability, thereby generating a standardized CBM production capacity feature set. The purpose of spatiotemporal alignment is to unify data from different sources into the same temporal and spatial framework, enabling cross-correlation and analysis.

[0022] For example, the specific spatiotemporal alignment process is as follows:

[0023] Step S121: extracting stratum dip parameters, fault distribution parameters, and fracture development parameters from the geological structure data to construct an initial geological feature set.

[0024] In this embodiment, key parameters are extracted from the geological structure data to construct an initial geological feature set. The formation dip parameter α can reflect the degree of inclination of the formation, which has an important influence on the migration and accumulation of coalbed methane. The dip values of the formation at different positions can be calculated by analyzing the seismic exploration data and combining the geological model. The fault distribution parameter β describes the distribution of faults in the target block, including the location, strike, length and other information of the faults. For example, it can be determined by combining seismic exploration data, geological mapping and drilling results. The fracture development parameter γ reflects the degree of development of fractures in the coal seam, such as the density and opening of the fractures. Relevant information can be obtained through microscopic observation of core samples, imaging logging and other methods.

[0025] By combining these extracted formation dip parameters α, fault distribution parameters β and fracture development parameters γ, an initial geological feature set G = {α, β, γ} is formed.

[0026] Step S122: extracting the porosity parameters, permeability parameters and gas content parameters from the reservoir physical property data to construct an initial reservoir feature set.

[0027] In this embodiment, key parameters are extracted from the reservoir physical property data to construct an initial reservoir feature set. The porosity parameter φ represents the ratio of the volume of pores in the coal seam to the total volume, which is an important indicator for measuring the coal seam's ability to store coalbed methane. The specific value can be determined by conducting porosity measurement experiments on core samples in the laboratory. The permeability parameter k reflects the ability of the coal seam to allow fluid to pass through, and is closely related to the flow of coalbed methane. Permeability is usually measured using a steady-state method or a non-steady-state method, and is obtained by simulating the flow of fluid in the core in the laboratory. The gas content parameter C represents the amount of coalbed methane contained in the coal seam, which can be estimated by directly measuring the gas content in the core sample, or by combining geological conditions and reservoir physical properties. The extracted porosity parameter φ, permeability parameter k, and gas content parameter C are combined to construct an initial reservoir feature set R={φ, k, C}.

[0028] Step S123: extracting the gas production sequence, pressure change sequence and production cycle stage labels from the historical production dynamic data to construct an initial dynamic feature set.

[0029] In this embodiment, key information is extracted from historical production dynamic data to construct an initial dynamic feature set. The gas production sequence Q(t) records coalbed methane production at different times t, reflecting the temporal variation of coalbed methane production. Gas production data at different time points can be obtained from the production monitoring system to form a time series. The pressure change sequence P(t) describes the temporal variation of pressure during coalbed methane production. Pressure changes affect the desorption and flow of coalbed methane. Similarly, a pressure sensor records pressure values at different times to construct a pressure change sequence. The production cycle stage label L is used to identify different stages of coalbed methane production, such as the initial gas production stage, the stable gas production stage, and the declining gas production stage. Based on the characteristics of the gas production sequence and the pressure change sequence, combined with production experience and professional knowledge, different production time periods can be divided into stages and assigned corresponding labels. The gas production sequence Q(t), the pressure change sequence P(t), and the production cycle stage label L are combined to construct the initial dynamic feature set D = {Q(t), P(t), L}.

[0030] Step S124: performing data preprocessing on the initial geological feature set, the initial reservoir feature set, and the initial dynamic feature set, wherein the data preprocessing includes missing value filling, outlier correction, and dimensional normalization.

[0031] In this embodiment, since the data in the initial geological feature set G, the initial reservoir feature set R, and the initial dynamic feature set D may contain missing values and abnormal values, and the dimensions of each feature may be different, data preprocessing is required.

[0032] There are many methods for filling missing values. For example, for continuous data (such as formation dip parameter α, porosity parameter φ, etc.), the mean filling method can be used, that is, calculating the mean of all non-missing values of the feature and using this mean to fill missing values; for discrete data (such as the production cycle stage label L), the mode filling method can be used, that is, finding the label value with the highest number of occurrences to fill missing values.

[0033] For outlier correction, a reasonable threshold range can be set to determine whether the data is an outlier. For example, for a gas production sequence Q(t), if the gas production at a certain point in time far exceeds the normal fluctuation range, this value is considered an outlier. Smoothing methods can be used to correct outliers, such as the moving average method, to replace the outlier with the average gas production of several time points before and after the time point.

[0034] Dimensional normalization is to eliminate the influence of different dimensions between features and make them have the same scale. You can use the minimum-maximum normalization method. For each feature, calculate its minimum value min and maximum value max. Then, normalize each feature value x using the formula (x-min) / (max-min) to map it to the interval [0, 1].

[0035] After the above data preprocessing operation, the preprocessed initial geological feature set G', initial reservoir feature set R' and initial dynamic feature set D' are obtained.

[0036] Step S125: perform time window alignment and spatial grid alignment on the pre-processed initial geological feature set, initial reservoir feature set and initial dynamic feature set to generate a standardized coalbed methane productivity feature set, wherein each feature unit in the standardized coalbed methane productivity feature set contains a unified timestamp identifier and spatial coordinate identifier.

[0037] In this embodiment, to achieve temporal and spatial unification of different feature sets, the preprocessed initial geological feature set G', initial reservoir feature set R', and initial dynamic feature set D' require time window alignment and spatial grid alignment. Time window alignment involves partitioning the temporal data within different feature sets into uniform time intervals to form a unified time window. For example, the gas production sequence Q(t), the pressure change sequence P(t), and time-related geological and reservoir feature data can be partitioned into one-day time windows, ensuring comparability within each time window. Spatial grid alignment involves dividing the target area into several spatial grids, each with a unique spatial coordinate identifier. The spatial information contained in the geological structure data and reservoir physical property data is matched to these grids, mapping the different feature data to the corresponding spatial grids. For example, geological features such as the formation dip parameter α and the fault distribution parameter β are assigned to corresponding spatial grids based on their spatial location. After time window alignment and spatial grid alignment, the three preprocessed feature sets are integrated to generate a standardized coalbed methane productivity feature set F. In this coalbed methane production capacity feature set, each feature unit contains a unified timestamp identifier t and spatial coordinate identifier (x, y), that is, F={F(t, x, y)}, where F(t, x, y) represents the feature vector at time t and spatial coordinate (x, y), which contains multiple feature information such as geology, reservoir and dynamics.

[0038] Step S130: extracting features from the standardized coalbed methane productivity feature set to generate a prediction feature set, wherein each feature unit in the prediction feature set includes dual coding of geological attribute features and dynamic attribute features.

[0039] In this example, to more effectively utilize the standardized coalbed methane production capacity feature set for production capacity prediction, feature extraction is performed to generate a prediction feature set that includes both geological and dynamic attribute features. The goal of feature extraction is to extract features that are crucial for production capacity prediction from the original feature set, thereby reducing the data dimensionality and improving prediction accuracy and efficiency.

[0040] For example, the specific feature extraction process is as follows:

[0041] Step S131: performing region segmentation processing on the initial geological feature set to generate a plurality of geological sub-region units.

[0042] In this embodiment, the initial geological feature set G' is subjected to regional segmentation processing to divide the target block into multiple geological sub-region units. The division can be performed based on the similarity of geological structures. For example, regions with similar stratum dip angles and similar fault distribution characteristics can be divided into the same sub-region. In specific operations, a cluster analysis method can be used, using geological characteristics such as stratum dip parameter α and fault distribution parameter β as clustering indicators to divide the space of the target block into different clusters, each cluster corresponding to a geological sub-region unit. For example, the set of geological sub-region units obtained by the division is set to S = {S1, S2, ..., Sn}, where n is the number of sub-regions.

[0043] Step S132: Synchronously call the reservoir physical property correlation analysis in each geological sub-region unit, combine the porosity parameters, permeability parameters and gas content parameters in the initial reservoir feature set, perform local feature extraction operations, and generate a geological sub-region feature set. The local feature extraction operations include formation continuity analysis, fault density calculation and fracture network topology modeling.

[0044] In this example, reservoir property correlation analysis and local feature extraction are performed within each geological subregion Si (i=1, 2, ..., n). First, the porosity parameter φ, permeability parameter k, and gas content parameter C in the initial reservoir feature set R' are combined to analyze their correlation with geological characteristics. For example, the correlation between porosity and formation dip, and the relationship between permeability and fault distribution, are studied. Formation continuity analysis analyzes formation information within the geological subregion to determine formation continuity. Formation continuity can be assessed based on factors such as the trend of formation dip and changes in formation thickness. Fault density calculation calculates the ratio of the number of faults within a geological subregion to the regional area, reflecting the density of faults within the region. Fracture network topology modeling models the fractures within the geological subregion and analyzes the connectivity and distribution patterns between fractures. Through these local feature extraction operations, a set of feature vectors is generated for each geological sub-region unit. These feature vectors are combined to form a geological sub-region feature set S'={S1', S2', ..., Sn'}, where Si' represents the feature vector of the i-th geological sub-region unit.

[0045] Step S133: performing cross-region feature aggregation processing on the geological sub-region feature set to generate a global geological feature set that integrates reservoir physical property parameters.

[0046] In this embodiment, cross-region feature aggregation is performed on the geological sub-region feature set S' to generate a global geological feature set that incorporates reservoir physical property parameters. Aggregation can be performed using a weighted average method, with weights determined based on factors such as the area and geological importance of each geological sub-region unit.

[0047] For example, larger subregions with larger geological characteristics that have a greater impact on productivity are assigned higher weights. Let the characteristic vector of the i-th geological subregion be Si', its weight be wi (i=1, 2, ..., n), and satisfy ∑wi=1. The global geological feature set G'' can then be calculated using the following formula: G''=∑(wi*Si'). This cross-regional feature aggregation process integrates the characteristic information of each geological subregion to produce a global geological feature set that reflects the geological characteristics of the entire target block, while also incorporating the influence of reservoir physical properties.

[0048] Step S134: performing time series decomposition processing on the initial dynamic feature set to generate a trend dynamic feature set and a periodic dynamic feature set, and associating the production cycle stage label as an auxiliary feature to the time series decomposition result.

[0049] In this embodiment, the initial dynamic feature set D' is subjected to time series decomposition to isolate its trend and periodic features. For the gas production sequence Q(t) and the pressure change sequence P(t), time series analysis methods, such as seasonal decomposition, can be employed. This method decomposes the time series into trend, seasonal, and residual components. The trend component reflects the long-term trend of the data over time, while the periodic component reflects the recurring pattern of data changes within a certain time period. Through time series decomposition, a trend dynamic feature set Qt and a periodic dynamic feature set Qp (for the gas production sequence), as well as Pt and Pp (for the pressure change sequence), are obtained. Simultaneously, the production cycle stage label L is associated with the time series decomposition results as an auxiliary feature. For example, the production cycle stage label at each time point is combined with the corresponding trend and periodic features to form a trend dynamic feature set Qt' and a periodic dynamic feature set Qp', as well as Pt' and Pp', with stage label information.

[0050] Step S135: generating a plurality of feature units according to the global geological feature set, the trend dynamic feature set and the periodic dynamic feature set, wherein the geological attribute features in each feature unit are synchronously associated with the initial reservoir feature set.

[0051] In this embodiment, multiple feature units are generated based on the global geological feature set G'', the trend-based dynamic feature set (Qt', Pt'), and the periodic dynamic feature set (Qp', Pp'). Each feature unit contains a dual encoding of geological attribute features and dynamic attribute features. When generating feature units, the geological attribute features in the global geological feature set are associated with the porosity parameter φ, permeability parameter k, and gas content parameter C in the initial reservoir feature set R', ensuring that the geological attribute features in each feature unit contain reservoir physical property information. At the same time, the trend-based dynamic features and periodic dynamic features are added to the feature units as dynamic attribute features. For example, for a certain time point t and spatial coordinates (x, y), the generated feature unit F(t, x, y) contains the geological attribute features (associated with reservoir physical properties) at that location and the corresponding trend-based and periodic dynamic features. These feature units are combined to form the prediction feature set P = {P(t, x, y)}, where each feature unit contains a dual encoding of geological attribute features and dynamic attribute features, which can be used for subsequent production capacity prediction.

[0052] Step S140: iteratively predicting the geological attribute characteristics and dynamic attribute characteristics based on the reinforcement learning strategy network to generate a target coalbed methane production capacity prediction result.

[0053] In this embodiment, a reinforcement learning strategy network is used to iteratively predict the geological and dynamic attributes in the prediction feature set P to generate a target coalbed methane production capacity forecast. The reinforcement learning strategy network can continuously adjust its strategy based on environmental feedback, thereby improving the accuracy of the forecast.

[0054] For example, the specific iterative prediction process is as follows:

[0055] Step S141: inputting the predicted feature set into a pre-trained reinforcement learning strategy network, wherein the reinforcement learning strategy network includes a feature selection module and a weight allocation module.

[0056] In this embodiment, the prediction feature set P is input into a pre-trained reinforcement learning strategy network. This network consists of a feature selection module and a weight assignment module. The function of the feature selection module is to filter out the most important features for capacity prediction from the prediction feature set, reducing the interference of unnecessary features on the prediction results. The weight assignment module assigns corresponding weights to each feature based on its importance to highlight the role of important features. The pre-trained reinforcement learning strategy network is trained on a large amount of historical data and has learned the relationship between geological attribute characteristics and dynamic attribute characteristics and coalbed methane capacity.

[0057] Step S142: The feature selection module is used to perform correlation scoring on the geological attribute features and the dynamic attribute features, and generate a feature importance score set that integrates the influence of reservoir physical properties.

[0058] In this embodiment, the feature selection module scores the correlation between geological attribute features and dynamic attribute features to generate a feature importance score set that integrates the influence of reservoir physical properties. The specific process is as follows:

[0059] For example, step S1421: perform dimension alignment processing on each feature unit in the geological attribute feature and the feature unit with the same timestamp identifier and spatial coordinate identifier in the dynamic attribute feature, and calculate mutual information through the decomposed single variable sequence to generate an initial correlation degree set.

[0060] In this embodiment, the geological attribute features and dynamic attribute features are first dimensionally aligned to ensure that feature units with the same timestamp identifier t and spatial coordinate identifier (x, y) can correspond one-to-one. Then, each feature unit is decomposed into a univariate sequence. For example, the multidimensional geological attribute feature vector and dynamic attribute feature vector are decomposed into multiple one-dimensional univariate sequences. Next, the mutual information between these univariate sequences is calculated. Mutual information is a measure of the correlation between two variables, reflecting the degree to which one variable contains information about the other variable. By calculating the mutual information between the univariate sequences of geological attribute features and dynamic attribute features, an initial correlation set I is obtained.

[0061] Step S1422: Based on the dual coding relationship between the geological attribute characteristics and the dynamic attribute characteristics, the time dimension and the spatial dimension of each feature unit in the initial association degree set are weighted by the attention mechanism to generate a spatiotemporal weighted score set.

[0062] In this embodiment, an attention mechanism is used to assign weights to the time and space dimensions of each feature unit in the initial relevance set I. The attention mechanism can automatically adjust the weights based on the importance of the features, highlighting key temporal and spatial information. Under the dual encoding relationship of geological attribute features and dynamic attribute features, considering that different time and spatial locations may have different impacts on coalbed methane production capacity, the attention mechanism assigns different weights to the time and space dimensions of each feature unit. For example, higher weights are assigned to feature units in key production stages and important geological areas. After weight assignment, a spatiotemporal weighted score set W is generated.

[0063] Step S1423: extracting the reservoir property influencing factor corresponding to each characteristic unit in the spatiotemporal weighted scoring set, wherein the reservoir property influencing factor is generated by the dynamic reservoir property parameters pre-associated in the geological attribute characteristics, and the dynamic reservoir property parameters are updated based on the reservoir property change data within the historical production cycle.

[0064] In this embodiment, the reservoir property influencing factor reflects the degree of influence of reservoir property parameters on coalbed methane production capacity prediction. Dynamic reservoir property parameters are pre-associated with the geological attribute features. These parameters are updated based on historical reservoir property change data over the production cycle. For example, parameters such as porosity, permeability, and gas content change with coalbed methane extraction. By analyzing historical production data, the changing patterns of these parameters can be determined and the dynamic reservoir property parameters can be updated. For each feature element in the spatiotemporal weighted scoring set W, a corresponding influencing factor f is extracted from the pre-associated dynamic reservoir property parameters. The calculation of the influencing factor f can be based on empirical models or machine learning algorithms, comprehensively considering the impact of changes in parameters such as porosity, permeability, and gas content on production capacity. For example, a multivariate linear regression model can be established with production capacity as the dependent variable and parameters such as porosity, permeability, and gas content as independent variables to obtain regression coefficients for each parameter. These regression coefficients can be used as part of the influencing factor. Furthermore, the influencing factor can be further adjusted by considering the changing trends and interactions between the parameters.

[0065] Step S1424: performing a product operation on each score value in the spatiotemporal weighted score set and the corresponding reservoir physical property influencing factor to generate an intermediate score set after physical property correction.

[0066] In this example, each score in the spatiotemporal weighted score set W is multiplied by the corresponding reservoir property impact factor f. Let the i-th score in the spatiotemporal weighted score set W be Wi, and the corresponding reservoir property impact factor be fi. Then, the i-th score Mi in the property-corrected intermediate score set M can be calculated using the formula Mi = Wi * fi. This multiplication incorporates the impact of reservoir properties into the score, allowing the score to more accurately reflect the importance of the feature to productivity prediction.

[0067] Step S1425: performing global normalization processing on the score values of all feature units in the intermediate score set, mapping them to a preset score dimension interval, and generating the feature importance score set.

[0068] In this embodiment, in order to make the feature importance scores comparable, it is necessary to perform global normalization on the intermediate score set M. A score dimension interval is preset, for example, [0, 1]. First, the minimum value minM and the maximum value maxM in the intermediate score set M are found. Then, for each score value Mi in the intermediate score set M, normalization is performed using the formula Si=(Mi-minM) / (maxM-minM), and it is mapped to the preset score dimension interval [0, 1]. After normalization, the feature importance score set S is obtained, where each score value Si represents the importance of the corresponding feature unit.

[0069] Step S143: dynamically assigning weights to each feature unit in the prediction feature set based on the feature importance score set by the weight assignment module to generate a weighted feature set.

[0070] In this embodiment, the weight allocation module dynamically assigns weights to each feature unit in the prediction feature set P based on the feature importance score set S. Let the j-th feature unit in the prediction feature set P be Pj, and the score value corresponding to Pj in the feature importance score set S be Sj. The weight wj assigned to Pj can be directly taken as Sj, that is, wj=Sj. Then, each feature unit Pj is multiplied by the corresponding weight wj to obtain the weighted feature unit Pj'=wj*Pj. All weighted feature units are combined to generate a weighted feature set P'={Pj'}.

[0071] Step S144: calling the regression prediction model to perform nonlinear mapping processing on the weighted feature set to generate an initial production capacity prediction result.

[0072] In this embodiment, a regression prediction model is called to perform nonlinear mapping processing on the weighted feature set P' to generate an initial capacity forecast result. The regression prediction model can be a neural network model, such as a multi-layer perceptron (MLP). The multi-layer perceptron consists of an input layer, a hidden layer, and an output layer. The input layer receives the feature vectors in the weighted feature set P', the hidden layer performs a nonlinear transformation on the input features, and the output layer outputs the capacity forecast value. During the training phase, the multi-layer perceptron is trained using a large amount of historical data, and the weights and biases of the model are adjusted so that the model can learn the nonlinear relationship between the weighted features and the capacity. During the prediction phase, the weighted feature set P' is input into the trained multi-layer perceptron, and after calculation and output by the model, the initial capacity forecast result Y0 is obtained.

[0073] Step S145: According to the error gradient between the initial production capacity prediction result and the actual gas production sequence, the parameter configurations of the feature selection module and the weight distribution module are synchronously adjusted to generate an optimized reinforcement learning strategy network.

[0074] In this embodiment, the error gradient between the initial capacity forecast result Y0 and the actual gas production sequence Q(t) is calculated. The error gradient reflects the degree of difference between the forecast and actual results, as well as the changing trend of this difference. The mean squared error (MSE) can be used as an error metric to calculate the average of the squared errors between the initial capacity forecast result Y0 and the actual gas production sequence Q(t) at each time point. The parameter configurations of the feature selection module and the weight assignment module are then adjusted synchronously based on the error gradient. For example, in the feature selection module, the threshold for selecting features can be adjusted to ensure that the selected features more accurately reflect their importance to the capacity forecast. In the weight assignment module, the weight assignment strategy can be adjusted to ensure that the weights of important features are more appropriate. By continuously adjusting the parameter configuration, the error is gradually reduced until the preset convergence conditions are met. After these adjustments, an optimized reinforcement learning policy network is generated.

[0075] Step S146: re-weighting the prediction feature set through the optimized reinforcement learning strategy network to generate a target coalbed methane production capacity prediction result.

[0076] In this embodiment, the prediction feature set P is re-input into the optimized reinforcement learning strategy network. The optimized feature selection module re-screens features that are more important for capacity prediction, and the weight assignment module re-weights feature units based on the updated feature importance scores. After re-weighting, a new weighted feature set P'' is obtained. The regression prediction model is then used to perform nonlinear mapping on the weighted feature set P'' to generate the target coalbed methane capacity prediction result Y.

[0077] Step S150: generating an optimization strategy set according to the error distribution between the target CBM production capacity prediction result and the historical production dynamic data, and feeding the optimization strategy set back to the CBM production control system to trigger parameter adjustment operation.

[0078] In this embodiment, in order to improve the accuracy of coalbed methane production capacity prediction and production efficiency, it is necessary to generate an optimization strategy set based on the error distribution between the target coalbed methane production capacity prediction result Y and the historical production dynamic data (mainly the actual gas production sequence Q(t)), and feed it back to the coalbed methane production control system to trigger the corresponding parameter adjustment operation.

[0079] For example, the specific process of generating an optimization strategy set is as follows:

[0080] Step S151: performing a confidence assessment on the target coalbed methane production capacity prediction result based on the historical production dynamic data to generate a prediction confidence score set.

[0081] In this embodiment, a confidence evaluation is performed on the target coalbed methane production capacity prediction result Y to generate a prediction confidence score set. The specific process is as follows:

[0082] Step S1511: extracting the predicted value fluctuation range of the target coalbed methane production capacity prediction result in different time windows to generate a time series fluctuation feature set.

[0083] In this example, the target CBM production capacity forecast result Y is divided into different time windows, for example, one week as a time window. For each time window, the maximum and minimum predicted values are calculated, and the difference between the two is the predicted value fluctuation range within that time window. The predicted value fluctuation ranges of all time windows are combined to generate a time series fluctuation feature set Vt.

[0084] Step S1512: Calculate the predicted value difference of the target coalbed methane production capacity prediction result within the spatial grid unit to generate a spatial difference feature set.

[0085] In this embodiment, the target block is divided into multiple spatial grid cells. For each spatial grid cell, the difference between the predicted values at different locations within the cell is calculated. The standard deviation can be used to measure the degree of difference in the predicted values. The predicted value differences of all spatial grid cells are combined to generate a spatial difference feature set Vs.

[0086] Step S1513: jointly analyze the temporal fluctuation feature set and the spatial difference feature set to determine a prediction stability score.

[0087] In this embodiment, a joint analysis is performed on the temporal fluctuation feature set Vt and the spatial difference feature set Vs. A weighted average approach can be used to assign different weights to temporal fluctuations and spatial differences based on their impact on predictive stability. For example, let the weight of the temporal fluctuation feature be wt, and the weight of the spatial difference feature be ws, where wt + ws = 1. For each time point and spatial location, the corresponding temporal fluctuation feature value and spatial difference feature value are weighted averaged to obtain the predictive stability score Sd.

[0088] Step S1514: Generate a prediction confidence score set based on the matching degree between the prediction stability score and the actual gas production fluctuation range in the historical production dynamic data, wherein each scoring unit in the prediction confidence score set includes a dual index of a timestamp identifier and a spatial coordinate identifier.

[0089] In this embodiment, the degree of match between the predicted stability score Sd and the actual gas production fluctuation range in the historical production dynamic data is calculated. The degree of match can be measured by calculating the similarity between the two, for example, using cosine similarity. For each time point and spatial position, the predicted stability score is compared with the actual gas production fluctuation range to obtain a matching value. These matching values are used as prediction confidence scores to generate a prediction confidence score set C. Each scoring unit in the set contains a dual index of the timestamp identifier t and the spatial coordinate identifier (x, y), that is, C = {C(t, x, y)}.

[0090] Step S152: determining a significant error feature subset based on the correlation analysis result between the prediction confidence score set and the error distribution.

[0091] In this embodiment, the correlation between the prediction confidence score set C and the error distribution is analyzed to determine the significant error feature subset. The specific process is as follows:

[0092] Step S1521: performing a binary cluster analysis on the error distribution based on a preset error threshold, and generating a target error region set having error values greater than the preset error threshold.

[0093] In this embodiment, a preset error threshold T is set. For each error value in the error distribution, a determination is made as to whether it exceeds the preset error threshold T. If so, the region corresponding to the error is marked as a target error region; otherwise, it is marked as a normal region. Through this binary cluster analysis, the error distribution is divided into target error regions and normal regions, generating a set of target error regions E.

[0094] Step S1522: extracting the original spatial coordinate identifiers corresponding to the target error region set, and matching the associated standardized spatial coordinate identifiers in the prediction confidence score set based on the spatiotemporal alignment mapping relationship.

[0095] In this embodiment, the original spatial coordinate identifier corresponding to each target error region is extracted from the target error region set E. Then, based on the previously performed spatiotemporal alignment mapping relationship, these original spatial coordinate identifiers are converted into the associated standardized spatial coordinate identifiers in the prediction confidence score set C. This allows the target error region to be mapped to the scoring unit in the prediction confidence score set.

[0096] Step S1523: Based on the standardized spatial coordinate identifiers with confidence scores lower than a preset threshold, trace back to the feature units in the predicted feature set that have been aligned with the spatial grid to generate a candidate feature subset.

[0097] In this embodiment, a preset confidence threshold Tc is set. For each scoring element in the prediction confidence score set C, a determination is made as to whether its score is below the preset threshold Tc. If so, the standardized spatial coordinate identifier corresponding to the scoring element is extracted. Then, based on these standardized spatial coordinate identifiers, the feature elements in the prediction feature set P that have been spatially aligned are traced back and these feature elements are combined to generate a candidate feature subset Fc.

[0098] Step S1524: Sort the candidate feature subsets based on the importance of feature contribution to generate a feature priority sequence.

[0099] In this embodiment, the feature contribution of each feature unit in the candidate feature subset Fc is calculated. The feature contribution can be measured by the feature importance score in the feature selection module. The feature units in the candidate feature subset Fc are sorted from large to small according to the feature contribution to generate a feature priority sequence Fp.

[0100] Step S1525: Generate a significant error feature subset based on the top N feature units in the feature priority sequence, where N is an integer dynamically adjusted according to the severity of the error distribution.

[0101] In this embodiment, the value of N is dynamically adjusted based on the severity of the error distribution. The more severe the error distribution, the larger the value of N, and the more feature units are selected. The top N feature units are selected from the feature priority sequence Fp and combined to generate the significant error feature subset Fs.

[0102] Step S153: extracting geological attribute features and dynamic attribute features corresponding to the significant error feature subset, and generating a feature optimization priority list.

[0103] In this example, corresponding geological and dynamic attribute features are extracted from the significant error feature subset Fs. These features are ranked according to their importance within the significant error feature subset to generate a feature optimization priority list Lp. Features in this list are those that have a significant impact on the production capacity forecast error and are therefore prioritized for optimization.

[0104] Step S154: Adjust the data acquisition equipment deployment plan based on the feature optimization priority list to generate a first optimization strategy.

[0105] In this embodiment, the data acquisition device deployment plan is adjusted according to the feature optimization priority list Lp to generate a first optimization strategy. The specific process is as follows:

[0106] Step S1541: Analyze the geological attribute features and dynamic attribute features in the feature optimization priority list to determine a set of key monitoring parameters.

[0107] In this example, the geological and dynamic attributes in the feature optimization priority list Lp are analyzed. Parameters closely related to data acquisition, such as formation dip, porosity, and gas production, are identified from these features. These parameters are combined to determine the key monitoring parameter set M.

[0108] Step S1542: adjusting the deployment position of the data acquisition equipment according to the spatial distribution density of the key monitoring parameter set, and generating an equipment position optimization plan.

[0109] In this embodiment, the spatial distribution density of each parameter in the key monitoring parameter set M is analyzed. In areas with high spatial distribution density, the number of deployed data acquisition devices is increased to obtain more detailed parameter information. In areas with low spatial distribution density, the number of data acquisition devices can be appropriately reduced. Based on this adjustment principle, the new deployment locations of the data acquisition devices are determined, and the device location optimization plan Dp is generated.

[0110] Step S1543: adjusting the sampling frequency of the data acquisition device according to the time variation frequency of the key monitoring parameter set, and generating a sampling frequency optimization solution.

[0111] In this embodiment, the time-varying frequency of each parameter in the key monitoring parameter set M is analyzed. For parameters with a faster time-varying frequency, the sampling frequency of the data acquisition device is increased to capture the rapid changes; for parameters with a slower time-varying frequency, the sampling frequency can be reduced. Based on this adjustment principle, a new sampling frequency for the data acquisition device is determined, generating a sampling frequency optimization solution Sp.

[0112] Step S1544: Integrate the device location optimization solution and the sampling frequency optimization solution to generate a first optimization strategy, wherein the first optimization strategy includes a configuration instruction set of device identification, location coordinates, sampling period and parameter type.

[0113] In this embodiment, the device location optimization solution Dp and the sampling frequency optimization solution Sp are integrated. Each data acquisition device is assigned a unique device identifier, which records its new location coordinates, sampling period, and the type of parameters to be monitored. This information is combined into a set of configuration instructions to generate the first optimization strategy O1.

[0114] Step S155: Dynamically configure the training frequency of the reinforcement learning strategy network according to the spatiotemporal variation pattern of the error distribution to generate a second optimization strategy.

[0115] In this embodiment, the training frequency of the reinforcement learning strategy network is dynamically configured according to the spatiotemporal variation pattern of the error distribution to generate the second optimization strategy. The specific process is as follows:

[0116] Step S1551: Decompose the error distribution in time dimension to generate long-term error trend characteristics and short-term error fluctuation characteristics.

[0117] In this embodiment, a time series analysis method is used to decompose the error distribution in the time dimension. The moving average method can be used to decompose the error distribution into a long-term trend component and a short-term fluctuation component. The long-term trend component reflects the trend of error changes over a long period of time, namely the long-term error trend characteristic Et; the short-term fluctuation component reflects the fluctuation of error over a short period of time, namely the short-term error fluctuation characteristic Es.

[0118] Step S1552: Decompose the error distribution into spatial dimensions to generate regional error aggregation features and discrete error dispersion features.

[0119] In this embodiment, the target block is divided into multiple spatial regions, and the distribution of errors in each region is analyzed. For regions with concentrated error distribution, regional error aggregation features Ea are defined; for regions with dispersed error distribution, discrete error dispersion features Ed are defined.

[0120] Step S1553: setting a basic training frequency according to the historical change rate of the long-term error trend characteristics and the regional error aggregation characteristics.

[0121] In this example, the historical rates of change of the long-term error trend feature Et and the regional error aggregation feature Ea are calculated. These historical rates reflect the rate of error change over a historical period. Based on these historical rates, the base training frequency Fb of the reinforcement learning policy network is set. The greater the historical rate of change, the higher the base training frequency, allowing for faster adjustment of model parameters to accommodate changes in error.

[0122] Step S1554: Calculate the dynamic adjustment amplitude according to the amplitude threshold of the short-term error fluctuation feature and the spatial density threshold of the discrete error dispersion feature.

[0123] In this embodiment, an amplitude threshold Ts for the short-term error fluctuation feature Es and a spatial density threshold Td for the discrete error dispersion feature Ed are set. When the amplitude of the short-term error fluctuation feature exceeds the amplitude threshold Ts, or the spatial density of the discrete error dispersion feature exceeds the spatial density threshold Td, the training frequency needs to be dynamically adjusted. The dynamic adjustment amplitude ΔF is calculated, and the dynamic adjustment amplitude can be calculated linearly or nonlinearly based on the degree to which the threshold is exceeded.

[0124] Step S1555: generating a training frequency configuration curve based on the linear superposition relationship between the basic training frequency and the dynamic adjustment amplitude.

[0125] In this embodiment, the basic training frequency Fb and the dynamic adjustment amplitude ΔF are linearly superimposed to obtain the training frequency F = Fb + ΔF. Based on the error distribution at different time points and spatial positions, the corresponding training frequencies are calculated and connected to generate the training frequency configuration curve Fc.

[0126] Step S1556: Convert the training frequency configuration curve into a parameter update instruction set of the reinforcement learning strategy network to generate a second optimization strategy.

[0127] In this embodiment, the training frequency configuration curve Fc is converted into a set of parameter update instructions for the reinforcement learning strategy network. Each parameter update instruction contains a timestamp, a training frequency, and information about the parameters to be updated. These parameter update instructions are combined to generate the second optimization strategy O2.

[0128] Step S156: Merge the first optimization strategy and the second optimization strategy to generate an optimization strategy set.

[0129] In this embodiment, the first optimization strategy O1 and the second optimization strategy O2 are merged to generate an optimization strategy set. The specific process is as follows:

[0130] For example, step S1561: obtaining the device location optimization scheme and the sampling frequency optimization scheme in the first optimization strategy, extracting the spatial coordinate identifier set in the device location optimization scheme and the time window identifier set in the sampling frequency optimization scheme.

[0131] In this embodiment, the device location optimization solution Dp and the sampling frequency optimization solution Sp are extracted from the first optimization strategy O1. The spatial coordinate identification set X of all data acquisition devices is extracted from the device location optimization solution Dp; and the sampling time window identification set T of each device is extracted from the sampling frequency optimization solution Sp.

[0132] Step S1562: Obtain the training frequency configuration curve in the second optimization strategy, and extract the time series parameters and the corresponding dynamic adjustment amplitude set in the training frequency configuration curve.

[0133] In this embodiment, the training frequency configuration curve Fc is obtained from the second optimization strategy O2. The time series parameter t and the corresponding dynamic adjustment amplitude set ΔF are extracted from the training frequency configuration curve Fc.

[0134] Step S1563: performing time dimension matching analysis on the time window identifier set and the time series parameters, and generating an association mapping table between the time-aligned device sampling period and the model training period.

[0135] In this embodiment, a matching analysis is performed on the time window identifier set T and the time series parameter t in the time dimension. First, the time range and resolution of the two sets are clarified and considered on the same time scale. For each time window in the time window identifier set T, the corresponding time point or time period in the time series parameter t is found. If the time window and the time series parameter have overlapping parts, they are considered to be matched. The information of the matched device sampling cycle and model training cycle is combined to generate an association mapping table M. The association mapping table records the correspondence between device sampling and model training at each time point or time period. For example, if there is a device sampling operation in a certain time window, and it corresponds to a training step in the model training cycle, this correspondence is recorded in the association mapping table.

[0136] Step S1564: Calculate the influence factor between the sampling interval in the sampling frequency optimization scheme and the frequency change rate in the dynamic adjustment amplitude set according to the period matching degree in the association mapping table.

[0137] In this embodiment, the influence factor between the sampling interval in the sampling frequency optimization scheme Sp and the frequency change rate in the dynamic adjustment amplitude set ΔF is calculated based on the period matching degree in the association mapping table M. First, determine the measurement method of the period matching degree, for example, use the ratio of the overlapping time of the sampling period and the training period to the total time to express it. For each set of matching relationships in the association mapping table M, analyze the relationship between the sampling interval and the frequency change rate. If there is a certain functional relationship between the sampling interval and the frequency change rate, the functional relationship is determined by statistical analysis of historical data or machine learning algorithm. The influence factor can be obtained by quantifying the functional relationship. For example, if it is found that the smaller the sampling interval, the greater the frequency change rate, an influence factor can be calculated based on this relationship to describe the strength of the interaction between them.

[0138] Step S1565: performing spatial density correction on the spatial coordinate identifier set in the device location optimization solution based on the influencing factors, and generating an optimization parameter configuration table that synchronously adapts the spatial coverage density and the temporal sampling frequency.

[0139] In this embodiment, the spatial coordinate identification set X in the device position optimization scheme Dp is corrected for spatial density using the influencing factors calculated previously. For example, the principle of spatial density correction is first clarified, that is, the distribution density of devices in different spatial positions is adjusted according to the influencing factors. If the influencing factors indicate that the interaction between the sampling interval and the frequency change rate will require more intensive sampling in certain areas, the number of data acquisition devices in these areas is increased; otherwise, the number of devices is reduced. Based on the adjusted device distribution and combined with the sampling frequency optimization scheme Sp, an optimization parameter configuration table P is generated that synchronously adapts the spatial coverage density and the time sampling frequency. The optimization parameter configuration table records the spatial coordinates, sampling frequency, and associated information with the model training of each device to ensure that spatial coverage and time sampling can be optimized collaboratively.

[0140] Step S1566: According to the parameter matching relationship in the optimization parameter configuration table, the device position optimization scheme, the sampling frequency optimization scheme and the training frequency configuration curve are jointly encoded to generate an optimization strategy set containing time-space unified dimension parameters, wherein the adjustment amount of each parameter unit in the optimization strategy set is synchronously updated based on the normalized time step and spatial grid ratio.

[0141] In this embodiment, the device location optimization scheme Dp, the sampling frequency optimization scheme Sp, and the training frequency configuration curve Fc are jointly encoded according to the parameter matching relationships in the optimization parameter configuration table P. The joint encoding process integrates the information from these three schemes into a unified representation, ensuring that each parameter has a clear time and space identifier. First, the time step and spatial grid are normalized to ensure that time and space have unified dimensions. For example, the time step is converted to a standard time unit and the spatial grid is divided according to a unified scale. Then, based on the matching relationships in the optimization parameter configuration table P, an adjustment value is determined for each parameter unit. The adjustment value is determined based on the normalized time step and spatial grid ratio to ensure that parameter adjustments are reasonable and consistent under different time and space conditions. The information of all parameter units is combined to generate an optimization strategy set O containing parameters with unified time and space dimensions. Each parameter unit in this optimization strategy set has a clear time and space identifier and adjustment value, which can accurately guide the coalbed methane production control system to perform parameter adjustments.

[0142] The generated optimization strategy set O is fed back to the coalbed methane production control system. Upon receiving the optimization strategy set, the control system adjusts the location and sampling frequency of the data acquisition equipment, as well as the training frequency of the reinforcement learning strategy network, based on the configuration instructions contained therein. For example, the installation location of the data acquisition equipment is adjusted according to the device location optimization plan, the sampling period of the equipment is set according to the sampling frequency optimization plan, and the training parameters of the reinforcement learning strategy network are updated according to the training frequency configuration curve. These adjustments can improve the accuracy of coalbed methane production capacity forecasts, optimize the coalbed methane production process, and enhance production efficiency and economic benefits.

[0143] Throughout the entire process, various technical measures were employed to protect data privacy and prevent data leakage. During the data collection phase, data collection equipment was encrypted to ensure data security during transmission. For example, collected data was encrypted using a symmetric encryption algorithm, allowing only authorized devices and systems to decrypt and process the data. For data storage, a secure database system was employed to implement access control and permission management. Only authorized personnel could access and manipulate data in the database. Furthermore, data was regularly backed up to prevent data loss. During the data processing and analysis phase, data was anonymized to remove sensitive information, such as specific geographic location and personal identity. For the training and use of artificial intelligence models, technologies such as federated learning were employed to avoid centralized data storage and processing, reducing the risk of data leakage. These privacy protection and leakage prevention technologies ensured the security and privacy of data during the coalbed methane production capacity forecasting process.

[0144] Through the above embodiments, by obtaining a multi-source coalbed methane production capacity data set, performing spatiotemporal alignment processing and feature extraction, using a reinforcement learning strategy network to perform production capacity prediction, and generating an optimization strategy set based on the error distribution between the prediction results and historical data, and feeding it back to the production control system for parameter adjustment, while adopting effective privacy protection and anti-leakage technical means, the accuracy of coalbed methane production capacity prediction and production efficiency can be improved.

[0145] Figure 2 A schematic diagram illustrates exemplary hardware and software components of an artificial intelligence-based coalbed methane production capacity prediction system 100, which can implement the concepts of the present application, as provided in some embodiments of the present application. For example, a processor 120 can be used in the artificial intelligence-based coalbed methane production capacity prediction system 100 to perform the functions described in the present application.

[0146] The AI-integrated coalbed methane production capacity prediction system 100 can be a general-purpose server or a special-purpose server, both of which can be used to implement the AI-integrated coalbed methane production capacity prediction method of this application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0147] For example, the coalbed methane production capacity prediction system 100 combined with artificial intelligence may include a network port 110 connected to the network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the coalbed methane production capacity prediction system 100 combined with artificial intelligence may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present application can be implemented according to these program instructions. The coalbed methane production capacity prediction system 100 combined with artificial intelligence also includes an I / O interface 150 between the computer and other input and output devices.

[0148] For ease of explanation, only one processor is described in the coalbed methane production capacity prediction system 100 combined with artificial intelligence. However, it should be noted that the coalbed methane production capacity prediction system 100 combined with artificial intelligence in the present application may also include multiple processors, so the steps performed by one processor described in the present application may also be performed jointly or individually by multiple processors. For example, if the processor of the coalbed methane production capacity prediction system 100 combined with artificial intelligence executes step A and step B, it should be understood that step A and step B may also be performed jointly by two different processors or individually in one processor. For example, the first processor executes step A, the second processor executes step B, or the first processor and the second processor execute steps A and B together.

[0149] In addition, an embodiment of the present invention also provides a readable storage medium, in which computer-executable instructions are preset. When the processor executes the computer-executable instructions, the coalbed methane production capacity prediction method combined with artificial intelligence as described above is implemented.

[0150] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, multiple features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. A coalbed methane production capacity prediction method combined with artificial intelligence, characterized in that: The method comprises: Acquire a multi-source coalbed methane production capacity data set for a target block, wherein the multi-source coalbed methane production capacity data set includes geological structure data, reservoir physical property data, and historical production performance data; Performing spatiotemporal alignment processing on the multi-source coalbed methane production capacity data set to generate a standardized coalbed methane production capacity feature set; Extracting features from the standardized coalbed methane productivity feature set to generate a prediction feature set, wherein each feature unit in the prediction feature set includes dual coding of geological attribute features and dynamic attribute features; Iteratively predict the geological attribute characteristics and dynamic attribute characteristics based on the reinforcement learning strategy network to generate a target coalbed methane production capacity prediction result; generating an optimization strategy set based on the error distribution between the target coalbed methane production capacity prediction result and the historical production dynamic data, and feeding the optimization strategy set back to the coalbed methane production control system to trigger a parameter adjustment operation; The performing of spatiotemporal alignment processing on the multi-source coalbed methane production capacity data set to generate a standardized coalbed methane production capacity feature set includes: Extracting formation dip parameters, fault distribution parameters, and fracture development parameters from the geological structure data to construct an initial geological feature set; Extracting porosity parameters, permeability parameters, and gas content parameters from the reservoir physical property data to construct an initial reservoir feature set; Extracting the gas production sequence, pressure change sequence and production cycle stage labels from the historical production dynamic data to construct an initial dynamic feature set; Performing data preprocessing on the initial geological feature set, the initial reservoir feature set, and the initial dynamic feature set, wherein the data preprocessing includes missing value filling, outlier correction, and dimensional normalization; Performing time window alignment and spatial grid alignment on the preprocessed initial geological feature set, initial reservoir feature set, and initial dynamic feature set to generate a standardized coalbed methane productivity feature set, wherein each feature unit in the standardized coalbed methane productivity feature set includes a unified timestamp identifier and spatial coordinate identifier; The feature extraction of the standardized coalbed methane production capacity feature set to generate a prediction feature set including geological attribute features and dynamic attribute features includes: Performing regional segmentation processing on the initial geological feature set to generate a plurality of geological sub-region units; Synchronously calling reservoir physical property correlation analysis in each geological sub-region unit, combining the porosity parameter, permeability parameter and gas content parameter in the initial reservoir feature set, performing a local feature extraction operation to generate a geological sub-region feature set, the local feature extraction operation including formation continuity analysis, fault density calculation and fracture network topology modeling; Performing cross-region feature aggregation processing on the geological sub-region feature set to generate a global geological feature set that integrates reservoir physical property parameters; Performing time series decomposition processing on the initial dynamic feature set to generate a trend dynamic feature set and a periodic dynamic feature set, and associating the production cycle stage label as an auxiliary feature with the time series decomposition result; Generating a plurality of feature units based on the global geological feature set, the trend dynamic feature set, and the periodic dynamic feature set, wherein the geological attribute features in each feature unit are synchronously associated with the initial reservoir feature set; The reinforcement learning strategy network is used to iteratively predict the geological attribute characteristics and dynamic attribute characteristics to generate a target coalbed methane production capacity prediction result, including: Inputting the predicted feature set into a pre-trained reinforcement learning strategy network, wherein the reinforcement learning strategy network includes a feature selection module and a weight allocation module; The feature selection module is used to score the correlation between the geological attribute features and the dynamic attribute features, and generate a feature importance score set that integrates the influence of reservoir physical properties; Dynamically assigning weights to each feature unit in the prediction feature set based on the feature importance score set by the weight assignment module to generate a weighted feature set; Calling the regression prediction model to perform nonlinear mapping processing on the weighted feature set to generate an initial capacity prediction result; According to the error gradient between the initial production capacity prediction result and the actual gas production sequence, the parameter configurations of the feature selection module and the weight allocation module are synchronously adjusted to generate an optimized reinforcement learning strategy network; The prediction feature set is re-weighted by the optimized reinforcement learning strategy network to generate a target coalbed methane production capacity prediction result.

2. The method for predicting coalbed methane production capacity combined with artificial intelligence according to claim 1, characterized in that: Generating an optimization strategy set according to the error distribution between the target coalbed methane production capacity prediction result and the historical production dynamic data includes: Performing a confidence assessment on the target coalbed methane production capacity prediction result based on the historical production dynamic data to generate a prediction confidence score set; Determining a significant error feature subset based on a correlation analysis result between the prediction confidence score set and the error distribution; Extracting geological attribute features and dynamic attribute features corresponding to the significant error feature subset and generating a feature optimization priority list; Adjusting the data acquisition device deployment plan based on the feature optimization priority list to generate a first optimization strategy; Dynamically configuring the training frequency of the reinforcement learning strategy network according to the spatiotemporal variation pattern of the error distribution to generate a second optimization strategy; The first optimization strategy and the second optimization strategy are combined to generate an optimization strategy set.

3. The method for predicting coalbed methane production capacity combined with artificial intelligence according to claim 2, characterized in that: The confidence evaluation of the target coalbed methane production capacity prediction result is performed based on the historical production dynamic data to generate a prediction confidence score set, including: Extracting the predicted value fluctuation range of the target coalbed methane production capacity prediction result in different time windows to generate a time series fluctuation feature set; Calculating the predicted value difference of the target coalbed methane production capacity prediction result within the spatial grid unit to generate a spatial difference feature set; Performing a joint analysis on the temporal fluctuation feature set and the spatial difference feature set to determine a prediction stability score; A prediction confidence score set is generated according to the matching degree between the prediction stability score and the actual gas production fluctuation range in the historical production dynamic data, wherein each scoring unit in the prediction confidence score set includes a dual index of a timestamp identifier and a spatial coordinate identifier.

4. The method for predicting coalbed methane production capacity combined with artificial intelligence according to claim 2, characterized in that: The determining of a significant error feature subset based on a correlation analysis result between the prediction confidence score set and the error distribution includes: Performing a binary cluster analysis on the error distribution based on a preset error threshold to generate a target error region set having an error value greater than the preset error threshold; Extracting the original spatial coordinate identifiers corresponding to the target error region set, and matching the associated standardized spatial coordinate identifiers in the prediction confidence score set based on a spatiotemporal alignment mapping relationship; Based on the standardized spatial coordinate identifiers whose confidence scores are lower than a preset threshold, tracing back to the feature units in the predicted feature set that have been aligned with the spatial grid, to generate a candidate feature subset; Sorting the candidate feature subsets based on the importance of feature contribution to generate a feature priority sequence; A significant error feature subset is generated based on the top N feature units in the feature priority sequence, where N is an integer dynamically adjusted according to the severity of the error distribution.

5. The method for predicting coalbed methane production capacity combined with artificial intelligence according to claim 2, characterized in that: The step of adjusting the data acquisition device deployment plan based on the feature optimization priority list to generate a first optimization strategy includes: Analyzing the geological attribute features and dynamic attribute features in the feature optimization priority list to determine a set of key monitoring parameters; Adjust the deployment location of the data acquisition equipment according to the spatial distribution density of the key monitoring parameter set to generate an equipment location optimization plan; Adjusting the sampling frequency of the data acquisition device according to the time-varying frequency of the key monitoring parameter set to generate a sampling frequency optimization solution; The device location optimization solution and the sampling frequency optimization solution are integrated to generate a first optimization strategy, wherein the first optimization strategy includes a configuration instruction set of device identification, location coordinates, sampling period and parameter type.

6. The method for predicting coalbed methane production capacity combined with artificial intelligence according to claim 2, characterized in that: The dynamically configuring the training frequency of the reinforcement learning strategy network according to the spatiotemporal variation pattern of the error distribution to generate a second optimization strategy includes: Decomposing the error distribution in a time dimension to generate long-term error trend characteristics and short-term error fluctuation characteristics; Decomposing the error distribution in spatial dimensions to generate regional error aggregation features and discrete error dispersion features; Setting a basic training frequency based on the historical change rate of the long-term error trend characteristics and the regional error aggregation characteristics; Calculating a dynamic adjustment amplitude based on an amplitude threshold of the short-term error fluctuation characteristic and a spatial density threshold of the discrete error dispersion characteristic; Generate a training frequency configuration curve based on the linear superposition relationship between the basic training frequency and the dynamic adjustment amplitude; The training frequency configuration curve is converted into a parameter update instruction set of a reinforcement learning strategy network to generate a second optimization strategy.

7. A coalbed methane production capacity prediction system combined with artificial intelligence, characterized in that: It includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the coalbed methane production capacity prediction method combined with artificial intelligence as described in any one of claims 1 to 6 above.

Images