Dynamic reservoir prediction method and device based on neural network

Abstract

The application discloses a dynamic reservoir prediction method and device based on a neural network. The method comprises the following steps: acquiring a seismic data body of a work area and a corresponding adjacent well logging curve of a well to be predicted, then, time-depth calibration and dynamic extraction of a well seismic trace are performed; before drilling, a training sample set and an initial prediction network model are established by using adjacent well seismic trace attribute data and adjacent well logging curves; during drilling, real-time well-seismic calibration is performed, a seismic data body of a well seismic trace of the well to be predicted is dynamically extracted, and a new training sample is constructed by combining the well logging curve, and an incremental learning model is fine-tuned; and the dynamic reservoir is updated by using the fine-tuned model. The application realizes a minute-level model response by means of a drilling data triggered sample update, model fine-tuning and prediction optimization automatic closed loop, guarantees adaptive optimization of a drilling trajectory, improves the dynamic reservoir prediction from post-correction to pre-judgment, improves a horizontal section reservoir drilling rate, and significantly reduces a geological risk.

Description

Technical Field

[0001] This application relates to the field of oilfield exploration and development technology, specifically to a dynamic reservoir prediction method, device, computer equipment, computer storage medium, and computer program product based on neural networks. Background Technology

[0002] In horizontal well drilling, the horizontal section of the reservoir often exhibits strong heterogeneity. The deviation between conventional reservoir prediction models and the actual drilled formation can be several meters. If only conventional prediction models are relied upon, the drill bit may accidentally exit the reservoir, resulting in well completion failure or a sharp drop in production.

[0003] Traditional horizontal well geological steering trajectory adjustment mainly relies on 90° phase shift prediction data volume. The prediction results have low vertical and horizontal resolution, and are not clear in terms of the pinch-out of horizontal sand bodies and the internal interlayers of sand bodies. While drilling-while-exploration data has a higher vertical resolution, the data lacks forward-looking information, which poses a significant challenge to on-site steering decisions. Summary of the Invention

[0004] In view of the above problems, this application is made in order to provide a method, apparatus, computer device, computer storage medium and computer program product based on neural networks to overcome the above problems or at least partially solve the above problems.

[0005] According to one aspect of the embodiments of this application, a dynamic reservoir prediction method based on neural networks is provided, the method comprising: Acquire seismic data volume of the work area and logging-while-drilling curves of adjacent wells corresponding to the well to be predicted; The seismic data volume of the work area and the logging-while-drilling curves of adjacent wells are subjected to time-depth calibration processing to obtain the first time-depth calibration result. Based on the first time-depth calibration result, the seismic data volume corresponding to the seismic traces of adjacent wells is dynamically extracted from the seismic data volume of the work area. A training sample set is constructed by sampling the logging-while-drilling curves of adjacent wells and the seismic data volume corresponding to the seismic traces of adjacent wells at a preset sampling interval. The logging-while-drilling data of adjacent wells is used as the label data, and the seismic data volume corresponding to the seismic traces of adjacent wells is used as the feature data. The initial prediction network model is obtained by training the preset neural network using the training sample set; The seismic data volume of the work area is input into the initial prediction network model to generate the initial reservoir prediction results. The initial reservoir prediction results are filtered based on a preset threshold, and the reservoir prediction model is generated based on the filtered initial reservoir prediction results. During the drilling process of the well to be predicted, at each preset drilling distance, the logging data of the well to be predicted within a preset range of the new well section is extracted. The seismic data volume of the working area and the logging data of the well to be predicted are subjected to time-depth calibration processing to obtain the second time-depth calibration result. Based on the second time-depth calibration result, the seismic data volume corresponding to the seismic trace of the well to be predicted is dynamically extracted from the seismic data volume of the working area. Using a preset sampling interval, the seismic data volume corresponding to the seismic trace of the well to be predicted and the logging data of the well to be predicted are sampled to construct new training samples. The training sample set is then updated by mixing the training samples with the old training samples in the training sample set based on a preset sample mixing ratio. The reservoir prediction model is fine-tuned based on the new training samples in the updated training sample set. The seismic data volume of the work area is then input into the fine-tuned reservoir prediction model to perform dynamic reservoir prediction and dynamically generate the target reservoir prediction results.

[0006] Furthermore, the method also includes: comparing the predicted boundary depth corresponding to the current reservoir prediction result with the predicted boundary depth corresponding to the previous reservoir prediction result, and outputting drilling commands based on the comparison results.

[0007] Furthermore, the drilling commands output based on the comparison results further include: If the comparison result shows that the depth deviation between the predicted boundary depth in the current reservoir prediction result and the predicted boundary depth in the previous reservoir prediction result is greater than a preset deviation threshold, an alarm message is output. Based on the alarm message, a trajectory adjustment command is output to adjust the trajectory and continue drilling. The alarm message includes: the predicted boundary depth, offset direction, and offset amount of the current reservoir prediction result. If the comparison result shows that the depth deviation between the predicted boundary depth in the current reservoir prediction result and the predicted boundary depth in the previous reservoir prediction result at the same location is less than or equal to a preset deviation threshold, then a continue drilling command is output to maintain the current trajectory and continue drilling.

[0008] Furthermore, the reservoir prediction model includes: an input layer, a convolutional layer, a long short-term memory network layer, and an output layer; Further fine-tuning of the reservoir prediction model based on new training samples in the updated training sample set includes: The weight parameters of the convolutional layer in the reservoir prediction model are frozen to keep the weight parameters of the convolutional layer unchanged. Using the new training samples in the updated training sample set, backpropagation and parameter updates are performed on the long short-term memory layer and output layer in the reservoir prediction model.

[0009] Furthermore, the seismic data volume for the work area includes: waveform indicator simulated seismic data volume; After acquiring the seismic data volume of the work area and the logging-while-drilling curves of adjacent wells corresponding to the well to be predicted, the method also includes: Bandpass filtering is applied to the waveform indication simulation data, and low-pass filtering is applied to the logging-while-drilling curves of adjacent wells.

[0010] Furthermore, the dynamic extraction of seismic data volumes corresponding to adjacent well seismic traces from the seismic data volume of the work area based on the first time-depth calibration results further includes: Using a triangulation algorithm, the spatial coordinates (X, Y, Z) of each sampling point are extracted sequentially along the well trajectory of adjacent wells, and the three seismic traces adjacent to each sampling point are determined. Based on the first time-depth calibration results, the seismic attribute values ​​of the three seismic traces were fitted to the corresponding positions of the well trajectories using resampling technology. The seismic attribute values ​​at the well point (X,Y) location are calculated using the inverse distance weighting method to obtain the seismic attribute data along the well trajectory. The seismic attribute data is the seismic data volume corresponding to the seismic traces of adjacent wells obtained by dynamic extraction.

[0011] According to another aspect of the embodiments of this application, a dynamic reservoir prediction device based on a neural network is provided, the device comprising: The acquisition module is suitable for acquiring seismic data volumes of the work area and logging-while-drilling curves of adjacent wells corresponding to the well to be predicted. The seismic data volume extraction module is suitable for performing time-depth calibration processing on the seismic data volume of the work area and the logging curves of adjacent wells to obtain the first time-depth calibration result. Based on the first time-depth calibration result, the seismic data volume corresponding to the seismic traces of adjacent wells is dynamically extracted from the seismic data volume of the work area. The training sample construction module is suitable for sampling the logging-while-drilling curves of adjacent wells and the seismic data volume corresponding to the seismic traces of adjacent wells at preset sampling intervals to construct a training sample set, wherein the logging-while-drilling data of adjacent wells is used as the label data and the seismic data volume corresponding to the seismic traces of adjacent wells is used as the feature data. The training module is suitable for training a pre-defined neural network using a training sample set to obtain an initial prediction network model; The reservoir prediction model generation module is suitable for inputting seismic data from the work area into the initial prediction network model, generating initial reservoir prediction results, filtering the initial reservoir prediction results based on a preset threshold, and generating a reservoir prediction model based on the filtered initial reservoir prediction results. The training sample update module is suitable for extracting logging data of the well to be predicted within a preset range of a new well section during the drilling process of the well to be predicted at each preset drilling distance. It performs time-depth calibration processing on the seismic data volume of the work area and the logging data of the well to be predicted to obtain a second time-depth calibration result. Based on the second time-depth calibration result, it dynamically extracts the seismic data volume corresponding to the seismic trace of the well to be predicted from the seismic data volume of the work area. It uses a preset sampling interval to sample the seismic data volume corresponding to the seismic trace of the well to be predicted and the logging data of the well to be predicted to construct new training samples. It then mixes the new training samples with the old training samples in the training sample set based on a preset sample mixing ratio to update the training sample set. The fine-tuning module is suitable for fine-tuning the reservoir prediction model based on new training samples in the updated training sample set. The reservoir dynamic prediction module is suitable for inputting seismic data from the work area into a finely tuned reservoir prediction model to perform dynamic reservoir prediction and dynamically generate target reservoir prediction results.

[0012] According to another aspect of the embodiments of this application, a computing device is provided, including: a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other through the communication bus; The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the above-mentioned neural network-based dynamic reservoir prediction method.

[0013] According to another aspect of the embodiments of this application, a computer storage medium is provided, which stores at least one executable instruction that causes a processor to perform operations corresponding to the above-described neural network-based dynamic reservoir prediction method.

[0014] According to another aspect of this application, a computer program product is provided, including at least one executable instruction that causes a processor to perform operations corresponding to the above-described neural network-based dynamic reservoir prediction method.

[0015] According to the technical solution provided in this application, adaptive nonlinear modeling is achieved through a neural network incremental learning framework. A convolutional-long short-term memory hybrid network is used to automatically extract the complex mapping relationship between seismic and well logging data, accurately capturing microscopic features such as thin reservoirs and pinch-out bodies. The automated closed loop of sample update → model fine-tuning → prediction optimization triggered by drilling data achieves minute-level model response, far exceeding the efficiency of manual adjustment. Three adjacent seismic traces near the adjacent well trajectory are obtained through triangulation. Inverse distance weighted interpolation is used to obtain the 90° phase-shifted seismic data volume near the well and the filtered waveform indicating the simulated seismic data volume. Real-time depth calibration is performed during drilling, matching the depth domain data of the drilling logging with the seismic time domain. Every 50m of drilling, the extraction range is expanded in real time, and seismic traces near the well are automatically extracted. Dynamic drilling logging data and real-time extracted seismic traces are used for automated sample updates, driving the dynamic update of reservoir prediction results. During the drilling process, combining drilling logging data and optimized reservoir prediction data, a higher resolution and accuracy reservoir prediction data volume is obtained using neural network methods, guiding horizontal well trajectory adjustment and drilling implementation. The neural network suppresses seismic noise and logging jitter through a feature weighting mechanism, reduces the false alarm rate, abandons experience dependence, and directly drives the guidance system with real-time prediction boundary (tolerance ±2m), ensuring adaptive optimization of drilling trajectory. Ultimately, it upgrades dynamic reservoir prediction from post-event correction to pre-event prediction, improves the drilling rate of horizontal reservoirs, and significantly reduces geological risks.

[0016] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the embodiments of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating a neural network-based dynamic reservoir prediction method according to an embodiment of this application is shown. Figure 2 A structural block diagram of a neural network-based dynamic reservoir prediction device according to an embodiment of this application is shown; Figure 3 A schematic diagram of the structure of a computing device according to an embodiment of this application is shown. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0019] Figure 1 A flowchart illustrating a neural network-based dynamic reservoir prediction method according to an embodiment of this application is shown, as follows: Figure 1 As shown, the method includes the following steps: Step S101: Obtain the seismic data volume of the work area and the logging-while-drilling curves of adjacent wells corresponding to the well to be predicted.

[0020] To better achieve reservoir prediction, a 90° phase-shift seismic data volume sensitive to geological body identification is selected, with waveform indicator simulation data used as the input for medium- and high-frequency data. Therefore, the seismic data volume for the work area mainly includes: a 90° phase-shift seismic data volume and a waveform indicator simulation seismic data volume, which remains unchanged throughout the prediction process.

[0021] This area can collect seismic data volumes for the work area and obtain logging-while-drilling curves from nearby drilled wells.

[0022] In one optional embodiment of this application, after acquiring the seismic data volume of the work area and the logging-while-drilling curves of adjacent wells corresponding to the well to be predicted, bandpass filtering can be performed on the waveform indication simulation data and low-pass filtering can be performed on the logging-while-drilling curves of adjacent wells.

[0023] For example, a 4th-order Butterworth filter is used to perform a 30–200Hz bandpass filter on the waveform indication simulated seismic data volume, resulting in the filtered waveform indication simulated seismic data volume. A 100th-order FIR filter is used to perform a 200Hz low-pass filter on the logging-while-drilling curves of adjacent wells for noise reduction, outputting the noise-reduced logging-while-drilling curves of the adjacent wells. Subsequent processing can then be based on the filtering results.

[0024] Step S102: Perform time-depth calibration processing on the seismic data volume of the work area and the logging-while-drilling curve of the adjacent well to obtain the first time-depth calibration result. Based on the first time-depth calibration result, dynamically extract the seismic data volume corresponding to the seismic trace of the adjacent well from the seismic data volume of the work area.

[0025] Specifically, the sonic transit time curve is extracted from the logging-while-drilling curve of the adjacent well, integrated, and the two-way travel time is calculated point by point. The logging data is then converted from the depth domain to the time domain, thereby establishing the initial time-depth relationship between the well and the seismic data.

[0026] Based on this, seismic traces near adjacent wells are extracted from the seismic data volume of the work area along the trajectory of adjacent wells. Polarity determination and spectral analysis are performed on the seismic traces near the wells to clarify the polarity and dominant frequency of the seismic data, and wavelets are selected based on this. At the same time, wave impedance and reflection coefficient sequences are calculated using the acoustic transit time and density information in the logging-while-drilling curves of the adjacent wells. The reflection coefficient sequences are then convolved with the selected wavelets to generate a synthetic seismic record.

[0027] Constrained by key stratigraphic interfaces in the region, the correspondence between the phase axes of the synthetic seismic record and the well-side seismic trace was repeatedly compared. When deviations occurred in wave group characteristics, the time-depth curves were locally adjusted using stretching or compression methods, supplemented by small time offsets if necessary, and the time-depth curves were finely corrected layer by layer. This process continued until the main reflection phase axis of the synthetic seismic record and the key phase axis of the well-side seismic trace were well aligned in waveform, phase, and energy, ultimately obtaining a precisely calibrated first time-depth calibration result and completing a reliable match between the logging-while-drilling curves of adjacent wells and the seismic data volume of the work area.

[0028] Based on the first time-depth calibration results obtained from the time-depth calibration process, the spatial range of seismic traces near adjacent wells is delineated. Depth stratification points within the wells are mapped to the seismic time domain, determining the selected time window and spatial location. Corresponding seismic traces are dynamically extracted from the seismic data volume of the work area, and then, through multi-trace overlay and weighted optimization, the seismic trace data volume near each adjacent well is finally generated. The seismic data volume corresponding to the seismic traces near adjacent wells changes dynamically with different well depths.

[0029] In one optional embodiment of this application, the seismic data volume corresponding to the seismic traces near adjacent wells can be dynamically extracted from the seismic data volume of the work area using the following method: A triangulation algorithm is used to sequentially extract the spatial coordinates (X, Y, Z) of each sampling point along the well trajectory of the adjacent wells, and the three seismic traces adjacent to each sampling point are determined; based on the first time-depth calibration result, the seismic attribute values ​​of the three seismic traces are fitted to the corresponding positions of the well trajectory using resampling technology; the seismic attribute values ​​at the (X, Y) position of the well point are calculated using the inverse distance weighting method to obtain the seismic attribute data along the well trajectory, wherein the seismic attribute data is the seismic data volume corresponding to the seismic traces near adjacent wells obtained by dynamic extraction.

[0030] Specifically, a triangulation algorithm is used to spatially discretize the well trajectory. Specifically, sampling points are extracted sequentially along the well trajectory at a preset sampling interval (e.g., 1 meter or an interval determined based on the seismic data sampling rate), resulting in a series of sampling points arranged in the order of the well trajectory. Each sampling point has precise spatial coordinates (X, Y, Z). Simultaneously, based on the spatial location of each sampling point, three adjacent seismic traces are retrieved and determined, and the original seismic data corresponding to these traces is read.

[0031] Based on the initial time-depth calibration results, three adjacent seismic traces were converted from the time domain to the depth domain, forming depth-domain seismic trace data. Then, for the actual depth value of each sampling point on the well trajectory, the attribute value at that depth was calculated using interpolation on the corresponding depth-domain seismic trace, thereby accurately fitting the seismic trace attribute values ​​to the depth position of the well trajectory. Finally, the fitted attribute values ​​of the three seismic traces corresponding to each sampling point were obtained.

[0032] For the fitted attribute values ​​of the three seismic traces, the inverse distance weighted method is used to calculate the final seismic attribute value at the current sampling point location. The planar distances from the three adjacent seismic traces to the sampling point are different; the closer the trace, the greater its influence on the attribute value of the sampling point, and therefore it is given a higher weight; the farther the trace, the smaller its weight. The weighted average of the three fitted attribute values ​​yields the seismic attribute value for that sampling point. By performing the above calculation sequentially according to the sampling points on the well trajectory, seismic attribute data along the entire well trajectory of adjacent wells can be dynamically extracted. This data is stored in the form of a seismic data volume, which can be used as input for subsequent reservoir prediction. The seismic data volume corresponding to the adjacent well-side seismic traces includes: a 90° phase shift seismic data volume and a waveform indicator simulated seismic data volume.

[0033] Step S103: Using a preset sampling interval, sample the logging-while-drilling curves of adjacent wells and the seismic data volumes corresponding to the seismic traces near adjacent wells to construct a training sample set, wherein the logging-while-drilling data of adjacent wells are used as label data and the seismic data volumes corresponding to the seismic traces near adjacent wells are used as feature data.

[0034] Using logging-while-drilling data from adjacent wells as label data and seismic data volumes corresponding to seismic traces near adjacent wells as feature data, both feature data and label data are sampled at preset intervals, such as 0.1 meters, and integrated along the well trajectory and other depths to generate a training sample set.

[0035] Step S104: Train the preset neural network using the training sample set to obtain an initial prediction network model.

[0036] The neural network structure is as follows: (1) Input layer: The input dimension is (batch_size, time_steps, features). time_steps corresponds to the number of sampling points in the depth direction (for example, for a depth range of 50m and a sampling interval of 0.1m, time_steps=500). features=2 (corresponding to two well-side seismic traces: 90° phase-shifted seismic data volume and 30~120Hz bandpass filtered waveform indicator simulated seismic data volume).

[0037] (2) Convolutional layer (Conv1D layer): filters=64: Uses 64 different convolution kernels, each kernel extracts a local spatial pattern of the input sequence (such as a segment of the shape of an earthquake waveform).

[0038] kernel_size=5: Each convolutional kernel covers features from 5 consecutive depth points at a time. This step is equivalent to automatically learning microscopic features such as edges and abrupt changes in seismic waveforms.

[0039] (3) LSTM layer (Long Short-Term Memory network layer): units=128: The LSTM has 128 elements and is used to capture long-range dependencies in seismic sequences (such as the gradual changes in reservoir depth). Each element contains a forget gate, an input gate, and an output gate to control the flow of information.

[0040] (4) Output layer: A fully connected layer is used to map the 128-dimensional feature vector output by the LSTM layer to a 1-dimensional output value, namely the predicted GR (natural gamma) value of the depth point.

[0041] The training process is as follows: (1) Data preprocessing: The feature data (seismic data volume) in the training sample set is Z-score standardized so that the mean is 0 and the standard deviation is 1.

[0042] (2) Data partitioning: After randomly shuffling the training samples according to depth order, the training set and the validation set are divided into a ratio of 8:2, that is, 80% is used for training and 20% is used for validation.

[0043] (3) Forward propagation and loss calculation: A batch of training samples is input into the network. The input data passes through the Conv1D layer, LSTM layer and fully connected layer in sequence to obtain the predicted GR value.

[0044] The Huber loss function is used to calculate the error between the predicted value and the true label (filtered GR curve). Huber loss is insensitive to outliers and is suitable for well logging data.

[0045] Its definition is: L_δ(y,f(x))={0.5 (yf(x))² for |yf(x)|≤δ; δ |yf(x)|-0.5δ²otherwise} In this task, δ = 1.0 is used.

[0046] Where x is the seismic data volume of the work area; y is the actual value of the reservoir; f(x) is the predicted value of the reservoir; δ is a hyperparameter; and L_δ is the loss function.

[0047] (4) Backpropagation and optimization: The Adam optimizer (adaptive moment estimator, learning rate lr=0.001) is used to calculate the gradient of the loss function with respect to the network weights and update the weights. Adam combines the momentum method with the adaptive learning rate mechanism, and has a faster convergence speed.

[0048] (5) Iteration and early stopping: Repeat the forward and backward propagation process for a total of 100 epochs. After each epoch, calculate the loss value using the validation set. If the validation loss does not decrease for 10 consecutive epochs, terminate the training early to prevent overfitting.

[0049] The initial prediction network model is output through training.

[0050] Step S105: Input the seismic data volume of the work area into the initial prediction network model to generate initial reservoir prediction results. Filter the initial reservoir prediction results based on a preset threshold and generate a reservoir prediction model based on the filtered initial reservoir prediction results.

[0051] The seismic data volume of the work area, including the 90° phase shift seismic data volume and the waveform indicator simulated seismic data volume, is spatially sampled using a uniform 25m × 25m horizontal grid (Inline × Xline). For each grid node, the seismic data volume (90° phase shift seismic data volume and waveform indicator simulated seismic data volume) of the well-side seismic trace at that node is extracted using the seismic trace extraction method described above. The seismic data volume of the well-side seismic trace at each node is input into a pre-trained initial prediction network model. This model internally learns and establishes a nonlinear mapping function from seismic attribute combinations to high-resolution gamma (GR) values. The GR value directly reflects the clay content, thus indicating reservoir (low GR) and non-reservoir (high GR). By inputting the seismic data volume of the work area into the model, a three-dimensional GR data volume with vertical resolution at the logging scale is output, enabling quantitative prediction from seismic response to reservoir parameters. The model outputs a set of predicted GR values ​​for each node with a vertical sampling interval of 0.5m; the prediction results of all nodes are stacked sequentially according to spatial coordinates to form a three-dimensional data volume (dimension: N_Inline × N_Xline × N_Depth / 0.5m), which is the initial GR prediction volume, and serves as the initial reservoir prediction result.

[0052] The data volume maintains a 25m×25m resolution in the horizontal direction, which matches the original CMP cell size or expected horizontal prediction accuracy of the 3D seismic data, ensuring the lateral resolution of the model while having a moderate computational load. It achieves an ultra-high resolution of 0.5m in the vertical direction, which is far superior to the vertical resolution of conventional seismic data (usually 2-4ms, about 5-10m). This density is the result of interpolation of the neural network model output, which is designed to match the sampling density of the well logging curve, enabling the model to finely characterize thin reservoirs with a thickness of less than 1m.

[0053] Based on this, reservoir screening is performed on the initial GR prediction volume according to a preset GR threshold (e.g., 65API, the specific value of which is determined by combining regional logging interpretation and core calibration): samples with GR values ​​below the GR threshold are identified as reservoirs, samples with GR values ​​above the GR threshold are classified as non-reservoirs, GR < 65API is sandstone, and GR > 65API is mudstone. Through this threshold screening, the continuous GR value volume is transformed into a binary, high-resolution reservoir spatial distribution. Subsequently, based on the screened reservoir spatial distribution, the final reservoir prediction model is generated.

[0054] In this application, the initial GR prediction volume is the global prediction result, while the logging-while-drilling curves of adjacent wells are local actual measurements. The logging-while-drilling curves (labels) of adjacent wells are the core data for training the model. The predicted values ​​of the initial GR prediction volume at the location of the adjacent well trajectory will highly match the actual logging curves (limited by model error). However, at locations far from the well point, the predicted values ​​are inferred by the model based on seismic characteristics. It can be said that the logging-while-drilling curves of adjacent wells are the standard answer learned by the model, while the GR prediction volume is the reasoning solution given by the model after applying the learned knowledge throughout the entire work area.

[0055] Step S106: During the drilling process of the well to be predicted, at each preset drilling distance, the logging data of the well to be predicted within a preset range of the new well section is extracted. The seismic data volume of the work area and the logging data of the well to be predicted are subjected to time-depth calibration. Based on the second time-depth calibration result, the seismic data volume corresponding to the seismic trace of the well to be predicted is dynamically extracted from the seismic data volume of the work area. Using a preset sampling interval, the seismic data volume corresponding to the seismic trace of the well to be predicted and the logging data of the well to be predicted are sampled to construct new training samples. The training sample set is updated by mixing the training samples with the old training samples in the training sample set based on a preset sample mixing ratio.

[0056] During the drilling process of the well to be predicted, the drilling depth will be monitored in real time. Every time a preset distance is drilled, for example, every 50m increase in drilling depth, an update process will be automatically triggered, triggering the execution of steps S106-S107.

[0057] Every 50m of drilling, logging data is extracted from the well to be predicted within ±10m of the new well section.

[0058] In this step, the time-depth calibration process and the process of extracting seismic data volumes from the well bypass to be predicted are similar to the corresponding process in step S102, and will not be described again here.

[0059] Then, the logging-while-drilling data of the well to be predicted is used as the label data, and the seismic data volume corresponding to the seismic trace next to the well to be predicted is used as the feature data. The feature data and label data are sampled at a preset sampling interval, for example, 0.1 meters, and the feature data and label data are integrated along the well trajectory and other depths to generate new training samples.

[0060] The new and old samples are mixed with a weight of 7:3 (that is, the new samples account for 70%) to update the training sample set, dynamically expand the sample library, and provide new data sources for model fine-tuning.

[0061] Step S107: Fine-tune the reservoir prediction model based on the new training samples in the updated training sample set, input the seismic data volume of the work area into the fine-tuned reservoir prediction model to perform dynamic reservoir prediction, and dynamically generate the target reservoir prediction result.

[0062] New training samples have been added to the training sample set. Here, the reservoir prediction model is fine-tuned using the new training samples from the updated training sample set. It should be noted that only the new samples are used here, and the old samples are not used, so as to avoid the model forgetting old knowledge (preventing overfitting to old samples).

[0063] It should be noted that during the drilling process of the well to be predicted, the seismic data volume corresponding to the seismic traces around the well changes dynamically. Therefore, the training sample set also changes dynamically. Here, the reservoir prediction model obtained after the previous fine-tuning is adjusted again based on the new training samples in the updated training sample set.

[0064] In one optional embodiment of this application, the reservoir prediction model includes: an input layer, a convolutional layer, a long short-term memory (LSTM) network layer, and an output layer; therefore, the fine-tuning of the reservoir prediction model based on new training samples in the updated training sample set further includes: freezing the weight parameters of the convolutional layer in the reservoir prediction model to keep the weight parameters of the convolutional layer unchanged; and using the new training samples in the updated training sample set to perform backpropagation and parameter update on the LSM layer and output layer in the reservoir prediction model.

[0065] Specifically, the full weights of the reservoir prediction model are loaded. The weight parameters of the convolutional layers in the model are set to non-trainable (i.e., frozen), thus preventing updates to these parameters during subsequent fine-tuning. Since the convolutional layers are mainly used to extract general local features from seismic waveforms, and these features are highly similar between new well sections and old samples, no significant adjustments are needed for new data.

[0066] Simultaneously, the LSTM layer and the final output layer in the model are kept in a trainable state. The LSTM layer is responsible for establishing long-range dependencies along the depth direction. New data may introduce new sedimentary cycles, thus requiring adaptive updates to the LSTM layer parameters using new data. The output layer is responsible for the final reservoir parameter mapping and is most sensitive to changes in data distribution, requiring the most adjustment based on information from new well sections. This approach of freezing the bottom layer and fine-tuning the top layer preserves learned general features while endowing the model with the ability to learn new sedimentary patterns.

[0067] The training sample set has been updated in step S106. Therefore, the incremental samples generated in step S106 are used here to fine-tune the model. That is, no old training samples are mixed at this time to avoid the model overfitting the old training samples during the fine-tuning process, which would cause the model to forget the old knowledge and ensure the model's comprehensive generalization ability on both new and old data.

[0068] Although it is a fine-tuned model, a validation set is set up here to validate the fine-tuned model. The validation set is divided from the new training samples. Specifically, the depth range corresponding to the new well section is divided proportionally along the depth direction, and the new training samples corresponding to the last 20% of the depth points with the largest depth are taken to form the validation set. The new training samples corresponding to the remaining depth points are used as the training sample set for this model fine-tuning to evaluate the model's prediction performance on the end data of the well section that was not involved in the training.

[0069] Fine-tuning training employed the Adam optimizer with a learning rate of 0.0001, approximately 1 / 10 of the initial training learning rate. The purpose of using a lower learning rate was to make small and stable parameter adjustments based on a well-converged model, preventing excessively large update strides from disrupting the general feature representations that are still propagated in the model after the convolutional layers are frozen.

[0070] This fine-tuning only involved training for 20 epochs. After each epoch, the Huber loss on the validation set was calculated, and its trend was monitored to ensure that the model did not overfit the limited amount of data from the new well section.

[0071] After training, the fine-tuned model weights are saved. Since only a few parameters of the LSTM and output layers are fine-tuned, and the input data is limited to samples within ±10m of the new well section, the computational cost of the entire model update is very small. In practice, it can usually be completed within 5 minutes, which can meet the needs of rapid iterative modeling in well sites.

[0072] After fine-tuning the reservoir prediction model, the seismic data volume of the work area is input into the fine-tuned reservoir prediction model for dynamic reservoir prediction, dynamically generating target reservoir prediction results. The model fine-tuning process is dynamic, and when the fine-tuned reservoir prediction model is applied to the seismic data volume again, an updated target reservoir prediction result is generated. By comparing the target reservoir prediction results before and after the update, changes in geological understanding revealed by the new data can be captured, thus achieving dynamic updating and prediction.

[0073] In one optional embodiment of this application, the method further includes: comparing the predicted boundary depth corresponding to the current reservoir prediction result with the predicted boundary depth corresponding to the previous reservoir prediction result; outputting a drilling command based on the comparison result; if the comparison result shows that the depth deviation between the predicted boundary depth in the current reservoir prediction result and the predicted boundary depth in the previous reservoir prediction result at the same location is greater than a preset deviation threshold (for example, the depth deviation between the predicted top boundary depth in the current reservoir prediction result and the predicted top boundary depth in the previous reservoir prediction result at the same location and / or the depth deviation between the predicted bottom boundary depth in the current reservoir prediction result and the predicted bottom boundary depth in the previous reservoir prediction result is greater than a preset deviation threshold), then an alarm message is output, and the method is further refined according to the alarm message. The system outputs a trajectory adjustment command to adjust the trajectory and continue drilling. The alarm information includes: the predicted boundary depth, offset direction, and offset amount of the current reservoir prediction result. If the comparison result shows that the depth deviation between the predicted boundary depth of the current reservoir prediction result and the predicted boundary depth of the previous reservoir prediction result at the same location is less than or equal to a preset deviation threshold (for example, the depth deviation between the predicted top boundary depth of the current reservoir prediction result and the predicted top boundary depth of the previous reservoir prediction result at the same location and / or the depth deviation between the predicted bottom boundary depth of the current reservoir prediction result and the predicted bottom boundary depth of the previous reservoir prediction result is less than or equal to the preset deviation threshold), then a continue drilling command is output to maintain the current trajectory and continue drilling.

[0074] This application acquires and monitors drilling depth in real time. Every 50 meters of cumulative drilling, the data update and intelligent prediction process is automatically initiated. The system retrieves seismic data volumes from the entire work area and the surrounding area of ​​the guide window, inputs them into the reservoir prediction model with fine-tuned parameters, and the model performs rapid inference calculations to output new three-dimensional natural gamma (GR) prediction results.

[0075] The system compares the reservoir top and bottom interface depths corresponding to the old and new GR predictions for the area in front of the drill bit (e.g., within 50 meters). When the depth difference between the predicted top or bottom boundary of the reservoir at the same plane coordinate (X,Y) exceeds a preset deviation threshold (e.g., 2 meters), the system immediately triggers an alarm. The alarm information includes: the new predicted interface depth, the offset direction (upward / downward), and the actual offset. Guiding engineers can combine this information with various drilling data such as imaging logging and gas logging data for comprehensive analysis, or they can directly accept the analysis conclusions output by the system.

[0076] Simultaneously, the system calculates the deviation between the current well inclination angle and the target reservoir center position in real time, and generates appropriate tool face angle and build-up rate control commands accordingly. For example, if the reservoir interface is predicted to shift upwards by 2 meters, a build-up command is issued to adjust the well inclination angle upwards. The relevant adjustment commands are transmitted to the rotary steerable tool via the drilling control system, which corrects the drill bit trajectory in real time, ensuring that the drill bit returns to and steadily drills along the center of the reservoir.

[0077] When the drill bit drills another 50 meters along the adjusted trajectory, the system repeats the data prediction, result comparison and deviation verification process to verify the effect of the trajectory adjustment. This constructs a fully automatic closed-loop drilling guidance system that combines prediction and judgment, command execution, real-time monitoring and iterative re-prediction.

[0078] It should be emphasized that this application is for dynamic reservoir prediction of the well to be predicted, and the training samples will be updated in the later stage based on the logging-while-drilling data of the well to be predicted.

[0079] The dynamic reservoir prediction method based on neural networks provided in this application utilizes an incremental learning framework for neural networks to adaptively model nonlinear processes. A convolutional-long short-term memory hybrid network automatically extracts the complex mapping relationship between seismic and well logging data, accurately capturing microscopic features such as thin reservoirs and pinch-out bodies. The automated closed loop of sample updates triggered by drilling data—model fine-tuning—prediction optimization—achieves minute-level model response, far exceeding the efficiency of manual adjustments. Three adjacent seismic traces near the well trajectory are obtained through triangulation. Inverse distance weighted interpolation yields the 90° phase-shifted seismic data volume near the well and the filtered waveform indicating the simulated seismic data volume. Real-time depth calibration is performed during drilling, matching the depth domain data of the drilling logging with the seismic time domain. Every 50m of drilling, the extraction range is expanded in real-time, automatically extracting seismic traces near the well. The dynamic drilling logging data and the real-time extracted seismic traces drive the dynamic updating of the reservoir prediction results. By combining logging-while-drilling data with optimized reservoir prediction data during the drilling process, a higher resolution and accuracy reservoir prediction data volume is obtained using neural network methods to guide horizontal well trajectory adjustments and drilling implementation. The neural network suppresses seismic noise and logging jitter through a feature weighting mechanism, reducing false alarm rates and abandoning reliance on experience. It directly drives the steering system with real-time prediction boundaries (tolerance ±2m), ensuring adaptive optimization of the drilling trajectory. Ultimately, dynamic reservoir prediction is upgraded from post-hoc correction to pre-hoc prediction, improving the reservoir encounter rate in horizontal sections and significantly reducing geological risks.

[0080] Figure 2 A structural block diagram of a neural network-based dynamic reservoir prediction device according to an embodiment of this application is shown, as follows: Figure 2 As shown, the device includes: The acquisition module 201 is suitable for acquiring the seismic data volume of the work area and the logging-while-drilling curves of adjacent wells corresponding to the well to be predicted; The seismic data volume extraction module 202 is adapted to perform time-depth calibration processing on the seismic data volume of the work area and the logging curve of the adjacent well to obtain a first time-depth calibration result, and dynamically extract the seismic data volume corresponding to the seismic trace of the adjacent well from the seismic data volume of the work area based on the first time-depth calibration result. The training sample construction module 203 is suitable for sampling the logging-while-drilling curves of adjacent wells and the seismic data volume corresponding to the seismic traces of adjacent wells using a preset sampling interval to construct a training sample set, wherein the logging-while-drilling data of adjacent wells is used as the label data and the seismic data volume corresponding to the seismic traces of adjacent wells is used as the feature data. Training module 204 is adapted to train a preset neural network using the training sample set to obtain an initial prediction network model; The reservoir prediction model generation module 205 is adapted to input the seismic data volume of the work area into the initial prediction network model, generate initial reservoir prediction results, filter the initial reservoir prediction results based on a preset threshold, and generate a reservoir prediction model based on the filtered initial reservoir prediction results. The training sample update module 206 is adapted to extract the logging data of the well to be predicted within a preset range of a new well section during the drilling process of the well to be predicted at each preset drilling distance, perform time-depth calibration processing on the seismic data volume of the work area and the logging data of the well to be predicted to obtain a second time-depth calibration result, dynamically extract the seismic data volume corresponding to the seismic trace of the well to be predicted from the seismic data volume of the work area based on the second time-depth calibration result, and use a preset sampling interval to sample the seismic data volume corresponding to the seismic trace of the well to be predicted and the logging data of the well to be predicted to construct new training samples, and mix them with the old training samples in the training sample set based on a preset sample mixing ratio to update the training sample set. The fine-tuning processing module 207 is adapted to fine-tune the reservoir prediction model based on new training samples in the updated training sample set. The reservoir dynamic prediction module 208 is suitable for inputting seismic data from the work area into the fine-tuned reservoir prediction model to perform dynamic reservoir prediction and dynamically generate the target reservoir prediction results.

[0081] Optionally, the device further includes: a comparison processing module, adapted to compare the predicted boundary depth corresponding to the current reservoir prediction result with the predicted boundary depth corresponding to the previous reservoir prediction result, and output drilling commands based on the comparison results.

[0082] Optionally, the comparison processing module is further adapted to: if the comparison result shows that the depth deviation between the predicted boundary depth in the current reservoir prediction result and the predicted boundary depth in the previous reservoir prediction result is greater than a preset deviation threshold, then an alarm message is output, and a trajectory adjustment command is output according to the alarm message to adjust the trajectory and continue drilling, wherein the alarm message includes: the predicted boundary depth, offset direction, and offset amount of the current reservoir prediction result; If the comparison result shows that the depth deviation between the predicted boundary depth in the current reservoir prediction result and the predicted boundary depth in the previous reservoir prediction result at the same location is less than or equal to a preset deviation threshold, then a continue drilling command is output to maintain the current trajectory and continue drilling.

[0083] Optionally, the reservoir prediction model includes: an input layer, a convolutional layer, a long short-term memory network layer, and an output layer; The fine-tuning module is further adapted to: freeze the weight parameters of the convolutional layer in the reservoir prediction model to keep the weight parameters of the convolutional layer unchanged; Using the new training samples in the updated training sample set, backpropagation and parameter updates are performed on the long short-term memory layer and output layer in the reservoir prediction model.

[0084] Optionally, the seismic data volume of the work area includes: waveform indication simulated seismic data volume; The device also includes a data preprocessing module, which is suitable for bandpass filtering of waveform indication simulation data and lowpass filtering of logging curves of adjacent wells.

[0085] Optionally, the seismic data volume extraction module is further adapted to: use a triangulation algorithm to sequentially extract the spatial coordinates (X,Y,Z) of each sampling point along the well trajectory of adjacent wells, and determine the three seismic traces adjacent to each sampling point; Based on the first time-depth calibration result, the seismic attribute values ​​of the three seismic traces are fitted to the corresponding positions of the well trajectory using resampling technology. The seismic attribute values ​​at the well point (X,Y) location are calculated using the inverse distance weighting method to obtain the seismic attribute data along the well trajectory. The seismic attribute data is the seismic data volume corresponding to the seismic traces of the adjacent wells, which are dynamically extracted.

[0086] The descriptions of the above modules refer to the corresponding descriptions in the method embodiments, and will not be repeated here.

[0087] The dynamic reservoir prediction device based on neural networks provided in this application utilizes an incremental learning framework for neural networks to adaptively model nonlinear processes. It automatically extracts the complex mapping relationship between seismic and well logging data through a convolutional-long short-term memory hybrid network, accurately capturing microscopic features such as thin reservoirs and pinch-out bodies. The automated closed loop of sample updates triggered by drilling data—model fine-tuning—prediction optimization—achieves minute-level model response, far exceeding the efficiency of manual adjustments. Three adjacent seismic traces near the well trajectory are obtained through triangulation. Inverse distance weighted interpolation yields the 90° phase-shifted seismic data volume near the well and the filtered waveform indicating the simulated seismic data volume. Real-time depth calibration is performed during drilling, matching the depth domain data of the drilling logging with the seismic time domain. Every 50m of drilling, the extraction range is expanded in real-time, automatically extracting seismic traces near the well. The dynamic drilling logging data and the real-time extracted seismic traces drive the dynamic updating of the reservoir prediction results. By combining logging-while-drilling data with optimized reservoir prediction data during the drilling process, a higher resolution and accuracy reservoir prediction data volume is obtained using neural network methods to guide horizontal well trajectory adjustments and drilling implementation. The neural network suppresses seismic noise and logging jitter through a feature weighting mechanism, reducing false alarm rates and abandoning reliance on experience. It directly drives the steering system with real-time prediction boundaries (tolerance ±2m), ensuring adaptive optimization of the drilling trajectory. Ultimately, dynamic reservoir prediction is upgraded from post-hoc correction to pre-hoc prediction, improving the reservoir encounter rate in horizontal sections and significantly reducing geological risks.

[0088] This application provides a non-volatile computer storage medium storing at least one executable instruction or computer program that enables a processor to perform the operation corresponding to the neural network-based dynamic reservoir prediction method in any of the above method embodiments.

[0089] This application provides a computer program product, which includes at least one executable instruction or computer program that enables a processor to perform the operation corresponding to the neural network-based dynamic reservoir prediction method in any of the above method embodiments.

[0090] Figure 3 The diagram shows a structural schematic of a computing device according to one embodiment of the present application. The specific embodiments of the present application do not limit the specific implementation of the computing device.

[0091] like Figure 3 As shown, the computing device may include: a processor 302, a communication interface 304, a memory 306, and a communication bus 308.

[0092] The processor 302, communication interface 304, and memory 306 communicate with each other via communication bus 308.

[0093] Communication interface 304 is used to communicate with other network elements such as clients or other servers.

[0094] The processor 302 is used to execute program 310, specifically to execute the relevant steps in the above-described embodiment of the dynamic reservoir prediction method based on neural networks.

[0095] Specifically, program 310 may include program code that includes computer operation instructions.

[0096] Processor 302 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application. The computing device includes one or more processors, which may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0097] Memory 306 is used to store program 310. Memory 306 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0098] Specifically, program 310 can be used to cause processor 302 to execute the neural network-based dynamic reservoir prediction method in any of the above method embodiments. The specific implementation of each step in program 310 can be found in the corresponding descriptions of the steps and units in the above embodiments of the neural network-based dynamic reservoir prediction method, and will not be repeated here. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the devices and modules described above can be referred to the corresponding process descriptions in the foregoing method embodiments, and will not be repeated here.

[0099] The algorithms and displays provided herein are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. The required structure for constructing such systems is apparent from the above description. Furthermore, this application is not directed to any particular programming language. It should be understood that the content of this application described herein can be implemented using various programming languages, and the above description of specific languages ​​is for the purpose of disclosing the best mode of implementation of this application.

[0100] Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of this application may be practiced without these specific details. In some instances, well-known methods, structures, and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0101] Similarly, it should be understood that, in order to streamline this disclosure and aid in understanding one or more of the various inventive aspects, in the above description of exemplary embodiments of this application, various features of this application are sometimes grouped together in a single embodiment, figure, or description thereof. However, this approach to disclosure should not be construed as reflecting an intention that the claimed application requires more features than expressly recited in each claim. Rather, as reflected in the claims, inventive aspects lie in fewer than all features of a single foregoing disclosed embodiment. Therefore, the claims following the detailed description are hereby expressly incorporated into that detailed description, wherein each claim itself is a separate embodiment of this application.

[0102] Those skilled in the art will understand that modules in the device of the embodiments can be adaptively changed and placed in one or more devices different from that embodiment. Modules, units, or components in the embodiments can be combined into a single module, unit, or component, and further, they can be divided into multiple sub-modules, sub-units, or sub-components. Except where at least some of such features and / or processes or units are mutually exclusive, any combination can be used to combine all features disclosed in this specification (including the accompanying claims, abstract, and drawings) and all processes or units of any method or device so disclosed. Unless expressly stated otherwise, each feature disclosed in this specification (including the accompanying claims, abstract, and drawings) may be replaced by an alternative feature that serves the same, equivalent, or similar purpose.

[0103] Furthermore, those skilled in the art will understand that although some embodiments described herein include certain features but not others included in other embodiments, combinations of features from different embodiments are intended to be within the scope of this application and form different embodiments. For example, in the claims, any one of the claimed embodiments can be used in any combination.

[0104] The various component embodiments of this application can be implemented in hardware, or as software modules running on one or more processors, or a combination thereof. Those skilled in the art will understand that microprocessors or digital signal processors (DSPs) can be used in practice to implement some or all of the functions of some or all of the components according to the embodiments of this application. This application can also be implemented as a device or apparatus program (e.g., a computer program and computer program product) for performing part or all of the methods described herein. Such an implementation of this application can be stored on a computer-readable medium, or can be in the form of one or more signals. Such signals can be downloaded from an Internet website, provided on a carrier signal, or provided in any other form.

[0105] It should be noted that the above embodiments are illustrative of this application and not restrictive, and that those skilled in the art can devise alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses should not be construed as limiting the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. This application can be implemented by means of hardware comprising several different elements and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, and third, etc., does not indicate any order. These words can be interpreted as names.

Claims

1. A dynamic reservoir prediction method based on neural networks, characterized in that, The method includes: Acquire seismic data volume of the work area and logging-while-drilling curves of adjacent wells corresponding to the well to be predicted; The seismic data volume of the work area and the logging-while-drilling curve of the adjacent well are subjected to time-depth calibration processing to obtain the first time-depth calibration result. Based on the first time-depth calibration result, the seismic data volume corresponding to the seismic trace of the adjacent well is dynamically extracted from the seismic data volume of the work area. A training sample set is constructed by sampling the logging-while-drilling curves of adjacent wells and the seismic data volume corresponding to the seismic traces of adjacent wells at a preset sampling interval. The logging-while-drilling data of adjacent wells is used as the label data, and the seismic data volume corresponding to the seismic traces of adjacent wells is used as the feature data. The preset neural network is trained using the training sample set to obtain an initial prediction network model; The seismic data volume of the work area is input into the initial prediction network model to generate initial reservoir prediction results. The initial reservoir prediction results are filtered based on a preset threshold, and a reservoir prediction model is generated based on the filtered initial reservoir prediction results. During the drilling process of the well to be predicted, at each preset drilling distance, the logging data of the well to be predicted within a preset range of the new well section is extracted. The seismic data volume of the work area and the logging data of the well to be predicted are subjected to time-depth calibration processing to obtain a second time-depth calibration result. Based on the second time-depth calibration result, the seismic data volume corresponding to the seismic trace of the well to be predicted is dynamically extracted from the seismic data volume of the work area. Using a preset sampling interval, the seismic data volume corresponding to the seismic trace of the well to be predicted and the logging data of the well to be predicted are sampled to construct new training samples. The training sample set is then updated by mixing the training samples with the old training samples in the training sample set based on a preset sample mixing ratio. The reservoir prediction model is fine-tuned based on the new training samples in the updated training sample set. The seismic data volume of the work area is then input into the fine-tuned reservoir prediction model to perform dynamic reservoir prediction and dynamically generate the target reservoir prediction results.

2. The method according to claim 1, characterized in that, The method further includes: comparing the predicted boundary depth corresponding to the current reservoir prediction result with the predicted boundary depth corresponding to the previous reservoir prediction result, and outputting drilling commands based on the comparison results.

3. The method according to claim 2, characterized in that, The step of outputting drilling commands based on the comparison results further includes: If the comparison result shows that the depth deviation between the predicted boundary depth in the current reservoir prediction result and the predicted boundary depth in the previous reservoir prediction result is greater than a preset deviation threshold, an alarm message is output. Based on the alarm message, a trajectory adjustment command is output to adjust the trajectory and continue drilling. The alarm message includes: the predicted boundary depth, offset direction, and offset amount of the current reservoir prediction result. If the comparison result shows that the depth deviation between the predicted boundary depth in the current reservoir prediction result and the predicted boundary depth in the previous reservoir prediction result at the same location is less than or equal to a preset deviation threshold, then a continue drilling command is output to maintain the current trajectory and continue drilling.

4. The method according to any one of claims 1-3, characterized in that, The reservoir prediction model includes: an input layer, a convolutional layer, a long short-term memory network layer, and an output layer; The step of fine-tuning the reservoir prediction model based on new training samples in the updated training sample set further includes: The weight parameters of the convolutional layer in the reservoir prediction model are frozen to keep the weight parameters of the convolutional layer unchanged. Using the new training samples in the updated training sample set, backpropagation and parameter updates are performed on the long short-term memory layer and output layer in the reservoir prediction model.

5. The method according to any one of claims 1-3, characterized in that, The seismic data body of the work area includes: waveform indication simulated seismic data body; After acquiring the seismic data volume of the work area and the logging-while-drilling curves of adjacent wells corresponding to the well to be predicted, the method further includes: Bandpass filtering is applied to the waveform indication simulation data, and low-pass filtering is applied to the logging-while-drilling curves of adjacent wells.

6. The method according to any one of claims 1-3, characterized in that, The step of dynamically extracting the seismic data volume corresponding to the seismic traces of adjacent wells from the seismic data volume of the work area based on the first time-depth calibration result further includes: Using a triangulation algorithm, the spatial coordinates (X, Y, Z) of each sampling point are extracted sequentially along the well trajectory of adjacent wells, and the three seismic traces adjacent to each sampling point are determined. Based on the first time-depth calibration result, the seismic attribute values ​​of the three seismic traces are fitted to the corresponding positions of the well trajectory using resampling technology. The seismic attribute values ​​at the well point (X,Y) location are calculated using the inverse distance weighting method to obtain the seismic attribute data along the well trajectory. The seismic attribute data is the seismic data volume corresponding to the seismic traces of the adjacent wells, which are dynamically extracted.

7. A dynamic reservoir prediction device based on neural networks, characterized in that, The device includes: The acquisition module is suitable for acquiring seismic data volumes of the work area and logging-while-drilling curves of adjacent wells corresponding to the well to be predicted. The seismic data volume extraction module is adapted to perform time-depth calibration processing on the seismic data volume of the work area and the logging-while-drilling curve of the adjacent well to obtain a first time-depth calibration result, and dynamically extract the seismic data volume corresponding to the seismic trace of the adjacent well from the seismic data volume of the work area based on the first time-depth calibration result. The training sample construction module is suitable for sampling the logging-while-drilling curves of adjacent wells and the seismic data volume corresponding to the seismic traces of adjacent wells at preset sampling intervals to construct a training sample set, wherein the logging-while-drilling data of adjacent wells is used as the label data and the seismic data volume corresponding to the seismic traces of adjacent wells is used as the feature data. The training module is adapted to train a preset neural network using the training sample set to obtain an initial prediction network model; The reservoir prediction model generation module is adapted to input the seismic data volume of the work area into the initial prediction network model, generate initial reservoir prediction results, filter the initial reservoir prediction results based on a preset threshold, and generate a reservoir prediction model based on the filtered initial reservoir prediction results. The training sample update module is suitable for extracting logging data of the well to be predicted within a preset range of a new well section during the drilling process of the well to be predicted at each preset drilling distance. It performs time-depth calibration processing on the seismic data body of the work area and the logging data of the well to be predicted to obtain a second time-depth calibration result. Based on the second time-depth calibration result, it dynamically extracts the seismic data body corresponding to the seismic trace of the well to be predicted from the seismic data body of the work area. Using a preset sampling interval, it samples the seismic data body corresponding to the seismic trace of the well to be predicted and the logging data of the well to be predicted to construct new training samples. Based on a preset sample mixing ratio, it mixes these samples with the old training samples in the training sample set to update the training sample set. The fine-tuning module is adapted to fine-tune the reservoir prediction model based on new training samples in the updated training sample set. The reservoir dynamic prediction module is suitable for inputting seismic data from the work area into a finely tuned reservoir prediction model to perform dynamic reservoir prediction and dynamically generate target reservoir prediction results.

8. A computer device, characterized in that, include: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction, which causes the processor to perform the operation corresponding to the neural network-based dynamic reservoir prediction method as described in any one of claims 1-6.

9. A computer storage medium, characterized in that, The computer storage medium stores at least one executable instruction, which causes the processor to perform the operation corresponding to the neural network-based dynamic reservoir prediction method as described in any one of claims 1-6.

10. A computer program product, characterized in that, It includes at least one executable instruction that causes the processor to perform the operation corresponding to the neural network-based dynamic reservoir prediction method as described in any one of claims 1-6.

Images