A pump pressure prediction method and apparatus

Abstract

The specification provides a pump pressure prediction method and device. According to the construction data, the initial feature sequence is determined; according to the initial feature sequence, the target feature sequence is determined; using the feature encoding module of the preset pump pressure prediction model, the target feature sequence is encoded and converted, and the high-dimensional code vector is determined; using the feature decoding module of the preset pump pressure prediction model, the high-dimensional code vector is processed with causal constraint correlation feature extraction, and the correlation feature vector is determined; using the routing unit of the preset pump pressure prediction model, the allocation weight for the multiple expert blocks is determined according to the correlation feature vector, and the multiple target expert blocks are determined from the multiple expert blocks according to the allocation weight; the output result of the target expert block is weighted and fused according to the allocation weight, and the decoding feature vector is determined; according to the decoding feature vector, the pump pressure prediction result in the target time period is determined. Thus, more accurate pump pressure prediction is realized.

Description

Technical Field

[0001] This specification belongs to the field of oil and gas development technology, and in particular relates to a pump pressure prediction method and device. Background Technology

[0002] With the industrial development of unconventional shale oil and gas resources, hydraulic fracturing technology has become a key means of reservoir stimulation. However, existing technologies mostly rely on human experience or post-event physical simulations, which have the problems of strong subjectivity and inability to accurately predict pump pressure based on multi-dimensional on-site construction data.

[0003] There is currently no effective solution to the above problems. Summary of the Invention

[0004] This specification provides a pump pressure prediction method and device, which solves the technical problems in the prior art where pump pressure prediction relies too much on human experience, resulting in strong subjectivity, and physical simulation methods have a lag, making it impossible to make real-time and accurate predictions based on multi-dimensional construction data on site.

[0005] This specification provides a pump pressure prediction method and apparatus, including: Acquire construction data from the target fracturing site, and determine an initial feature sequence based on the construction data; Based on the initial feature sequence, the distribution feature parameters within the corresponding preset sliding window are determined, and the initial feature sequence is normalized using the distribution feature parameters to determine the target feature sequence. The target feature sequence is encoded and converted using the feature encoding module of the preset pump pressure prediction model to determine a high-dimensional encoding vector; wherein the preset pump pressure prediction model includes at least: a feature encoding module, a feature decoding module, a routing unit, and multiple expert blocks; Using the feature decoding module of the preset pump pressure prediction model, the high-dimensional encoded vector is subjected to correlation feature extraction processing with causal constraints to determine the correlation feature vector; Using the routing unit of the preset pump pressure prediction model, the allocation weights for the multiple expert blocks are determined according to the associated feature vectors, and multiple target expert blocks are determined from the multiple expert blocks according to the allocation weights. The routing unit of the preset pump pressure prediction model performs weighted fusion processing on the output of the target expert block according to the allocated weights to determine the decoding feature vector; wherein, the output of the target expert block is the result obtained by processing the associated feature vector using the target expert block; Based on the decoded feature vector, the pump pressure prediction result within the target time period is determined through dimension mapping processing. In one embodiment, determining the distribution feature parameters within a corresponding preset sliding window based on the initial feature sequence, and normalizing the initial feature sequence using the distribution feature parameters to determine the target feature sequence, includes: Using a preset sliding window, the initial feature sequence is divided into multiple sub-feature sequences; where each sub-feature sequence corresponds to one time step. Based on each sub-feature sequence, the mean and standard deviation of the features at each time step are calculated and used as the distribution feature parameters at each time step. By utilizing the distribution characteristic parameters of each time step, the sub-feature sequences of the corresponding time steps are normalized to obtain the target feature sequence.

[0006] In one embodiment, the preset pump pressure prediction model is a fusion model of LSTM-MoE. The feature encoding module of the preset pump pressure prediction model performs encoding transformation processing on the target feature sequence to determine a high-dimensional encoding vector, including: Using the first layer of the long short-term memory neural network in the feature encoding module, the first-level feature extraction is performed on the target feature sequence to determine the first encoded feature vector with a first preset dimension; Using the second layer of long short-term memory neural network in the feature encoding module, a second-level feature extraction is performed on the first encoded feature vector to determine the high-dimensional encoded vector with a second preset dimension; Wherein, the second preset dimension is greater than the first preset dimension.

[0007] In one embodiment, the feature decoding module utilizing a preset pump pressure prediction model performs causal constraint-based correlation feature extraction processing on the high-dimensional encoded vector to determine the correlation feature vector, including: Using the normalization layer in the feature decoding module, a normalized feature vector is determined based on the high-dimensional encoded vector; The self-attention layer in the feature decoding module is used to perform feature association calculation on the normalized feature vector to determine the temporal correlation feature vector; wherein, the self-attention layer is provided with a temporal causal mask; The associated feature vector is determined by summing the time-related feature vector and the high-dimensional encoding vector using a preset residual connection structure.

[0008] In one embodiment, the routing unit utilizing a preset pump pressure prediction model determines allocation weights for the plurality of expert blocks based on the associated feature vectors, and determines a plurality of target expert blocks from the plurality of expert blocks based on the allocation weights, including: Using the routing unit, feature mapping processing is performed on the associated feature vector to determine the initial weight values ​​for each expert block; The initial weight values ​​are normalized using a preset activation function to determine the assigned weights for each expert block. The numerical relationships of the assigned weights are filtered to determine multiple target expert blocks from the multiple expert blocks.

[0009] In one embodiment, the routing unit utilizing a preset pump pressure prediction model performs weighted fusion processing on the output of the target expert block according to the assigned weights to determine the decoded feature vector, including: Based on the associated feature vector, feature transformation processing is performed on each of the target expert blocks to determine the output result of each target expert block; Using the routing unit, the output results of each target expert block and the assigned weights are weighted and fused to determine the hybrid expert feature vector; Based on the hybrid expert feature vector and the associated feature vector, the decoding feature vector is determined using a preset residual connection structure.

[0010] In one embodiment, the preset pump pressure prediction model further includes a linear layer, and the step of determining the pump pressure prediction result within the target time period based on the decoded feature vector through dimensionality mapping processing includes: Using the linear layer in the preset pump pressure prediction model, the dimension mapping process is performed on the decoded feature vector to determine the predicted feature vector; Based on the first preset duration corresponding to the predicted feature vector, the target time period relative to the current time is determined; The pump pressure prediction result within the target time period is determined by mapping the numerical points contained in the predicted feature vector to each sampling time within the target time period.

[0011] In one embodiment, the method further includes: Based on the pump pressure prediction results, determine the abnormal construction conditions at the fracturing construction site; Based on the aforementioned abnormal construction conditions, determine the corresponding risk warning information; Based on the risk warning information, a strategy for adjusting the construction plan at the fracturing site is determined.

[0012] This specification provides a pump pressure prediction device, including: The initial sequence determination module is used to acquire construction data at the target fracturing construction site and determine the initial feature sequence based on the construction data. The target sequence determination module is used to determine the distribution feature parameters within a corresponding preset sliding window based on the initial feature sequence, and to normalize the initial feature sequence using the distribution feature parameters to determine the target feature sequence. The first vector determination module is used to encode and transform the target feature sequence using the feature encoding module of the preset pump pressure prediction model to determine the high-dimensional encoded vector; wherein, the preset pump pressure prediction model includes at least: a feature encoding module, a feature decoding module, a routing unit, and multiple expert blocks; The second vector determination module is used to perform correlation feature extraction processing with causal constraints on the high-dimensional encoded vector by using the feature decoding module of the preset pump pressure prediction model to determine the correlation feature vector. The expert block determination module is used to determine the allocation weights for the multiple expert blocks based on the associated feature vectors using the routing unit of the preset pump pressure prediction model, and to determine multiple target expert blocks from the multiple expert blocks based on the allocation weights. The third vector determination module is used to utilize the routing unit of the preset pump pressure prediction model to perform weighted fusion processing on the output result of the target expert block according to the allocated weights, and determine the decoded feature vector; wherein, the output result of the target expert block is the result obtained by processing the associated feature vector using the target expert block; The result determination module is used to determine the pump pressure prediction result within the target time period based on the decoded feature vector through dimension mapping processing.

[0013] This specification also provides an electronic device including a processor and a memory for storing processor-executable instructions, wherein the processor implements a pump pressure prediction method when executing the instructions.

[0014] This specification also provides a computer-readable storage medium having computer instructions stored thereon, which, when executed, implement a pump pressure prediction method.

[0015] Based on the pump pressure prediction method provided in this specification, construction data of the target fracturing construction site is acquired, and an initial feature sequence is determined based on the construction data. Based on the initial feature sequence, distribution feature parameters within a corresponding preset sliding window are determined, and the initial feature sequence is normalized using the distribution feature parameters to determine the target feature sequence. The target feature sequence is encoded and converted using the feature encoding module of a preset pump pressure prediction model to determine a high-dimensional encoding vector. The preset pump pressure prediction model includes at least: a feature encoding module, a feature decoding module, a routing unit, and multiple expert blocks. The high-dimensional feature sequence is encoded and converted using the feature decoding module of the preset pump pressure prediction model. The encoded vector undergoes causal constraint-based correlation feature extraction processing to determine the correlation feature vector. Using the routing unit of a pre-defined pump pressure prediction model, weights are assigned to the multiple expert blocks based on the correlation feature vector, and multiple target expert blocks are identified from the multiple expert blocks based on the assigned weights. The routing unit of the pre-defined pump pressure prediction model then performs weighted fusion processing on the output results of the target expert blocks based on the assigned weights to determine the decoded feature vector. The output result of the target expert block is the result obtained by processing the correlation feature vector using the target expert block. Based on the decoded feature vector, the pump pressure prediction result within the target time period is determined through dimension mapping processing. In this way, by acquiring construction data from the target fracturing construction site and performing normalization processing using distribution feature parameters, the subjectivity problem caused by reliance on human experience in existing technologies is overcome. By utilizing the feature encoding module, the feature decoding module with causal constraint-based correlation feature extraction processing, and the routing unit's weighted fusion processing of the output results of multiple target expert blocks, in-depth mining and dynamic scheduling of high-dimensional encoded vectors and correlation feature vectors in multi-dimensional construction data from the site are achieved. This solves the problem that physical simulation methods cannot make real-time predictions for multi-dimensional data, thereby achieving more accurate pump pressure prediction. Attached Figure Description

[0016] To more clearly illustrate the embodiments of this specification, the accompanying drawings used in the embodiments will be briefly introduced below. The drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic flowchart of a pump pressure prediction method provided in one embodiment of this specification; Figure 2 This is a schematic diagram of the electronic device structure provided in one embodiment of this specification; Figure 3 This is a schematic diagram of the structural composition of a pump pressure prediction device provided in one embodiment of this specification; Figure 4 This is a schematic diagram of the overall architecture of a pump pressure prediction method provided in one embodiment of this specification. Detailed Implementation

[0018] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0019] See Figure 1 As shown in the embodiments of this specification, a pump pressure prediction method is provided, which is specifically applied to the server side. In specific implementation, the method may include the following: S101: Obtain construction data from the target fracturing site and determine an initial feature sequence based on the construction data; S102: Based on the initial feature sequence, determine the distribution feature parameters within the corresponding preset sliding window, and use the distribution feature parameters to normalize the initial feature sequence to determine the target feature sequence; S103: Using the feature encoding module of the preset pump pressure prediction model, the target feature sequence is encoded and converted to determine a high-dimensional encoding vector; wherein, the preset pump pressure prediction model includes at least: a feature encoding module, a feature decoding module, a routing unit, and multiple expert blocks; S104: Using the feature decoding module of the preset pump pressure prediction model, perform correlation feature extraction processing with causal constraints on the high-dimensional encoded vector to determine the correlation feature vector; S105: Using the routing unit of the preset pump pressure prediction model, the associated feature vector is used to determine the allocation weight for the multiple expert blocks, and multiple target expert blocks are determined from the multiple expert blocks according to the allocation weight. S106: The routing unit of the preset pump pressure prediction model performs weighted fusion processing on the output result of the target expert block according to the allocated weights to determine the decoding feature vector; wherein, the output result of the target expert block is the result obtained by processing the associated feature vector using the target expert block; S107: Based on the decoded feature vector, determine the pump pressure prediction result within the target time period through dimension mapping processing.

[0020] The aforementioned construction data can refer to the raw physical signals collected in real time by sensors at the fracturing operation site. These typically include key parameters such as displacement, oil pressure, sand ratio, and total fluid injection volume, serving as the raw input information for the entire prediction system.

[0021] The aforementioned initial feature sequence can refer to a structured data sequence formed by cleaning, aligning, and time-segmenting the original construction data and arranging them in chronological order. It transforms scattered physical signals into initial feature expressions that the model can process.

[0022] The aforementioned preset sliding window can refer to a time span with a fixed length that is pre-set when processing time series, used to delineate the data range for calculating local statistical features, ensuring that the model can capture dynamic changes within a specific time domain.

[0023] The aforementioned distribution characteristic parameters can refer to statistical indicators that reflect the distribution characteristics of data, such as the mean and standard deviation, calculated within a preset sliding window. These parameters are used to characterize the central tendency and dispersion of construction data within the current window.

[0024] The aforementioned target feature sequence can refer to the standard numerical sequence obtained by normalizing the initial feature sequence using distribution feature parameters (such as Z-score transformation), which aims to eliminate the differences between different physical dimensions and improve the convergence speed and stability of the model.

[0025] The aforementioned preset pump pressure prediction model can be a deep learning model that integrates a long short-term memory neural network (LSTM) and a hybrid expert (MoE) mechanism, responsible for performing the nonlinear mapping from construction data to prediction results.

[0026] The aforementioned feature encoding module can refer to the unit in the model responsible for preliminary feature extraction. It captures the long-term and short-term dependencies in time-series signals by processing the target feature sequence and transforms physical features into hidden semantic expressions.

[0027] The aforementioned high-dimensional encoding vector can refer to the hidden layer state vector output by the feature encoding module. By increasing the feature dimension, it maps the original temporal information to a high-dimensional abstract space, providing a rich information carrier for subsequent deep decoding.

[0028] The aforementioned feature decoding module can refer to the unit in the model responsible for high-level association extraction. Its core lies in processing high-dimensional encoded vectors through a mechanism with causal masking, aiming to uncover the deep coupling relationship between features while adhering to the chronological logic.

[0029] The aforementioned associated feature vector can refer to the feature vector produced by the feature decoding module under the condition of satisfying causal constraints. It accurately describes the intrinsic correlation strength between the historical construction status and the current and future prediction targets.

[0030] The aforementioned routing unit can refer to the decision center in a hybrid expert architecture (MoE), which is responsible for dynamically calculating the contribution of each expert block based on the input associated feature vector and determining the allocation scheme of computing resources.

[0031] The aforementioned expert block can refer to multiple independent sub-network units running in parallel within the model. Each expert block spontaneously learns the physical laws under specific working conditions (such as the steady period or the fluctuating period) during training, thereby achieving specialized processing of complex features.

[0032] The aforementioned weight allocation can refer to the probability value or score calculated by the routing unit for each expert block, used to quantify the importance or trust level of different expert blocks when processing the current feature.

[0033] The aforementioned target expert block can refer to a specific group of expert units selected from multiple candidate expert blocks based on the size relationship of the assigned weights, which participates in the actual calculation, reflecting the model's sparse selection and optimization of the calculation path.

[0034] The aforementioned decoding feature vector can refer to the comprehensive vector generated by linearly weighting and fusing the processing results of multiple target expert blocks according to the assigned weights, representing the final decision features of the model decoding stage.

[0035] The aforementioned dimensional mapping process refers to the operation of projecting high-dimensional decoded feature vectors onto a low-dimensional numerical space through a fully connected linear transformation, thereby realizing the transformation from abstract features to specific physical values ​​(pump pressure).

[0036] The pump pressure prediction results mentioned above can refer to the expected pressure values ​​or evolution curves generated after dimensional mapping processing for a target time period.

[0037] In some embodiments, determining the initial feature sequence based on the construction data may specifically include: Based on the construction data, perform time axis alignment processing on the multi-source data to determine the synchronous construction data; Based on the synchronous construction data, the effective time period is segmented using the preset construction node information to determine the target construction segment data; Based on the target construction section data, sequence encapsulation processing is performed in chronological order to determine the initial feature sequence.

[0038] Specifically, during the initial processing of construction data from the target fracturing site, the diversity of on-site sensors—such as instantaneous displacement gauges, pressure sensors, and sand mixers—often results in different sampling frequencies, and even clock asynchrony. By performing time axis alignment processing on this multi-source data, linear interpolation or resampling algorithms can be used to map sensor signals of different frequencies onto a unified time scale, thereby obtaining synchronous construction data containing complete dimensions. This processing method ensures that within the same time segment, parameters such as displacement, oil pressure, and sand ratio have a strict spatial correspondence.

[0039] After acquiring synchronous construction data, to eliminate interference from invalid periods such as equipment debugging and standby shutdowns on model predictions, pre-set construction node information—such as sand pump start signals, oil pressure exceeding threshold times, or construction command markers—is used to precisely scan the data stream. By identifying the core operational interval from the start of sand injection to the construction depressurization time, effective time period segmentation is performed, thereby accurately locking onto the target construction segment data. This process filters out random noise in non-construction states, allowing the model to focus on the true dynamic data reflecting downhole fracture propagation.

[0040] For the extracted target construction segment data, it is arranged in an ordered manner according to the time step, and the numerical vectors of each physical dimension on the time axis are tensorized and stacked. Through this sequence encapsulation process, the scattered construction records are transformed into an initial feature sequence with temporal logical attributes. This initial feature sequence not only preserves the physical causal chain of pressure response caused by displacement fluctuations during fracturing, but also provides a structured data foundation for subsequent statistical feature extraction within the sliding window.

[0041] By aligning the construction data with the time axis and segmenting it into effective time periods, data pollution caused by heterogeneous sensor sampling and non-operational conditions is eliminated. This ensures that the initial feature sequences entering the model have extremely high physical consistency and data purity, providing high-quality training and inference samples for more accurate pump pressure prediction from the source.

[0042] In some embodiments, this specification includes the following in practice: S1: Data acquisition and preprocessing logic.

[0043] In the construction data acquisition stage, real-time physical parameters during actual fracturing operations are collected through sensors, including but not limited to construction oil pressure, instantaneous discharge rate, and construction sand ratio, to ensure that the sample set covers historical working conditions under various geological conditions.

[0044] In the data cleaning and preprocessing stage, the raw data is segmented into effective time periods using pre-defined construction node information. The core operational interval from the start to the end of sand addition is automatically extracted to eliminate noise interference from the pre- and post-construction phases. Simultaneously, outlier correction is performed on negative numbers, abnormal jumps, and other values ​​that significantly deviate from physical logic in the raw data. Subsequently, the processed data is normalized to eliminate dimensional differences between different dimensions of data, improving the model's convergence efficiency.

[0045] The normalization process specifically includes: setting the initial size of the preset sliding window and adjusting the span of the sliding window as the sampling points move; calculating the mean and standard deviation and other distribution characteristic parameters within each window; and extracting only the end data corresponding to the preset window size within the extended window for normalization calculation, that is, subtracting the corresponding distribution characteristic parameters from the data at each time step before performing division scaling.

[0046] S2: Model architecture design.

[0047] The preset pump pressure prediction model adopts an LSTM-MoE fusion architecture, and the specific construction steps are as follows: Feature encoding module: Features are extracted from the target feature sequence using a two-layer Long Short-Term Memory (LSTM) neural network. The first LSTM has a hidden layer dimension of 256, and the second LSTM has a hidden layer dimension of 512. Through two consecutive extraction layers, the initial data is mapped to a high-dimensional encoding vector of 600×512.

[0048] Feature decoding module: It consists of 60 cascaded decoding blocks. Each decoding block integrates a normalization layer (LayerNorm), a self-attention layer with a temporal causality mask, a second normalization layer, and a hybrid expert (MoE) layer in sequence.

[0049] Connection structure: Residual connection structures are set before the normalization layer and after the self-attention layer, as well as before the second normalization layer and after the MoE layer.

[0050] Dimension Mapping Layer: A linear layer is deployed at the end of the model to map the decoded feature vectors into prediction vectors of length 300, thus enabling the output of pump pressure prediction results for the next 5 minutes. The total number of trainable parameters for the entire machine is close to 200 million.

[0051] S3: Model training strategy.

[0052] The large model training process includes: loading a sample set from a structured data file, constructing a feature mapping relationship using an extended sliding window method to ensure that every 600 seconds of input data corresponds to 300 seconds of target prediction data, and performing dynamic normalization as described above.

[0053] A random sampling data loader is constructed to split the dataset into training and validation sets at an 8:2 ratio. The Adam optimizer is used for parameter optimization, with an initial learning rate of 0.0001, a fixed batch size of 32, and an epoch of 100. During training, the weight parameters, loss function curve, and learning rate evolution are recorded in real-time for each epoch. For deep network architectures, gradient pruning is introduced to suppress gradient explosion during training.

[0054] S4: On-site deployment and application.

[0055] During the on-site deployment phase, a computing platform equipped with high-performance computing terminals (such as NVIDIA RTX 5080 or higher-performance graphics cards) is used to deploy the trained model by calling the optimal model weights. The system accesses real-time construction raw data transmitted from the fracturing site, performs real-time inference through the model, generates predicted pump pressure curves, and provides high-precision decision support for construction safety monitoring.

[0056] Based on the above embodiments, the beneficial effects of this specification compared with the prior art are as follows: First, this specification achieves deep fitting and pattern learning for complex construction conditions. It directly acquires real construction data from the target fracturing site (including hydraulic pressure, instantaneous displacement, and sand ratio). Compared to simulation data generated by traditional numerical simulations or physical calculations, the initial feature sequence determined by this method can more realistically reproduce the nonlinear dynamic changes during the operation, making the final pump pressure prediction results more consistent with engineering reality and exhibiting stronger generalization ability. Simultaneously, the multi-dimensional raw data contains richer hidden information, helping the model to more comprehensively learn the intrinsic mechanism of pump pressure changes during construction.

[0057] Secondly, this specification ensures a high degree of consistency between training and inference logic through an extended sliding window mechanism. This specification employs a normalization method based on an extended sliding window, effectively resolving the prediction bias caused by inconsistent calculation benchmarks in traditional normalization during the training phase (relying on global parameters) and the inference phase (only obtaining current parameters). By dynamically determining the distribution characteristic parameters (mean and standard deviation) within the corresponding preset sliding window, and performing normalization only on the data at the end of the window to determine the target feature sequence, this not only ensures the mathematical alignment of the training and inference processes but also fully utilizes the causal characteristics of time series data, maximizing the preservation of background distribution information and minimizing the negative impact of normalization operations on prediction accuracy.

[0058] Third, this specification significantly improves feature representation capabilities and computational efficiency by utilizing an LSTM-MoE fusion architecture. The pre-built pump prediction model constructed in this specification greatly reduces the intervention of manual feature engineering through the synergistic effect of the feature encoding module (two-layer LSTM) and the feature decoding module (including MoE layers), and is not limited by dimensional space. Compared with traditional machine learning or simple RNN models, this architecture, with a larger trainable parameter space, achieves a more efficient feature mining mode through correlation feature extraction processing with causal constraints. By dynamically calling multiple expert blocks through the routing unit, both the accuracy of the prediction results and the reliability of the model are taken into account.

[0059] Fourth, this manual represents a technological leap from "post-event analysis" to "pre-event prediction." It overcomes the limitations of traditional physical models that rely on post-event simulation. By deploying a trained pump pressure prediction model on-site, it receives construction data in real time and rapidly outputs pump pressure prediction results for the target time period, proactively identifying potential construction anomalies and triggering risk warnings. This provides a scientific basis for dynamically optimizing construction plans and adjusting strategies, offering significant engineering guidance and constructive contributions to ensuring fracturing safety, preventing inter-well interference, and improving oilfield recovery.

[0060] In some embodiments, the method further includes the following: determining the distribution feature parameters within a corresponding preset sliding window based on the initial feature sequence, and normalizing the initial feature sequence using the distribution feature parameters to determine the target feature sequence. S1: Using a preset sliding window, the initial feature sequence is divided into multiple sub-feature sequences; where each sub-feature sequence corresponds to one time step; S2: Based on each sub-feature sequence, calculate the mean and standard deviation of the features at each time step, which are used as the distribution feature parameters at each time step; S3: Using the distribution characteristic parameters of each time step, normalize the sub-feature sequences of the corresponding time steps to obtain the target feature sequence.

[0061] Specifically, when processing the initial feature sequence, a preset sliding window of length L (e.g., 300 or 600 sampling points) is selected. As the construction time step progresses, this window slides sequentially along the time axis, dividing the continuous feature flow into a series of sub-feature sequences with overlapping regions. Each specific time step t is associated with a unique sub-feature sequence, which fully preserves the local historical dynamics prior to the current prediction node, establishing the time boundary for the model's feature perception.

[0062] For each sub-feature sequence, statistical analysis is performed on its physical dimensions, such as displacement, pressure, and sand ratio. By traversing all data points within the window, the arithmetic mean and standard deviation of each dimension at that specific time step are calculated. These distributional characteristic parameters accurately capture the center of numerical distribution and the intensity of fluctuations in fracturing operations within local time periods, quantitatively abstracting the stable and rapid phases of the construction process, so that the background distribution of each time step has an independent and accurate parameter description.

[0063] Using the distribution characteristic parameters calculated at each time step, numerical mapping is performed on the sub-feature sequences within the current window. By subtracting the mean of the corresponding time step from the original values ​​and dividing by the standard deviation, the input data with different physical dimensions and numerical spans are mapped to a dimensionless standard normal distribution space, thus producing the target feature sequence. This mapping mechanism ensures that the data entering subsequent encoding modules maintains a consistent numerical magnitude while highlighting the evolution trend of the current data relative to the local background at that moment.

[0064] By introducing a dynamic normalization mechanism based on a sliding window, the problem of distribution drift of pump pressure data caused by non-stationarity in fracturing operations is solved. This allows the model to ignore the absolute magnitude difference of sensor signals and instead focus on the deep-seated patterns of feature changes, significantly improving the convergence accuracy and prediction stability of the model under multiple operating conditions.

[0065] In some embodiments, the preset pump pressure prediction model is a fusion model of LSTM-MoE. The feature encoding module of the preset pump pressure prediction model performs encoding transformation processing on the target feature sequence to determine a high-dimensional encoding vector. In specific implementations, the method may also include the following: S1: Using the first layer of long short-term memory neural network in the feature encoding module, perform first-level feature extraction on the target feature sequence to determine a first encoded feature vector with a first preset dimension; S2: Using the second layer of long short-term memory neural network in the feature encoding module, perform second-level feature extraction on the first encoded feature vector to determine the high-dimensional encoded vector with a second preset dimension; Wherein, the second preset dimension is greater than the first preset dimension.

[0066] The aforementioned preset pump pressure prediction model adopts an LSTM-MoE (Long Short-Term Memory Mixture of Experts) fusion large model architecture. It should be clarified that LSTM stands for Long Short-Term Memory Neural Network, primarily used for extracting temporal features from construction data; MoE stands for Hybrid Expert Model (or Hybrid Expert System), used to achieve sparse activation and expert-based collaborative computation under complex working conditions.

[0067] Specifically, when performing encoding transformation on the target feature sequence, the first layer of the Long Short-Term Memory (LSTM) neural network deployed at the bottom of the feature encoding module is used to initially perceive the normalized construction features of the input. This layer of the neural network uses an internal forget gate, input gate, and output gate loop mechanism to perform nonlinear feature extraction on the changing patterns of parameters such as displacement and pressure on the time axis, capturing the short-term fluctuation correlations between construction parameters. Through this layer's processing, the low-dimensional physical features of the input are mapped to a latent space with a first preset dimension d_1 (e.g., 256 dimensions), forming the first encoded feature vector, thus initially completing the transformation from raw numerical values ​​to temporal semantic features.

[0068] The first encoded feature vector is passed as input to the second layer of the Long Short-Term Memory (LSTM) neural network. This layer, based on the temporal features extracted in the previous layer, performs deeper feature deconstruction and spatial transformation to determine the high-dimensional encoded vector. In this process, by setting the number of hidden layer nodes in the neural network to a second preset dimension d_2 (e.g., 512 dimensions), a significant expansion of the feature representation space is achieved. Since d_2 > d_1, this spatial dimensionality-upgrading effectively solves the problem of overlapping physical information and excessive coupling in the low-dimensional feature space, providing the model with a more ample parameter space to decouple complex construction conditions, such as distinguishing between normal displacement adjustment signals and abnormal formation pressure precursors.

[0069] The first-level feature extraction focuses on the initial perception and short-range correlation modeling of the original time-series signal, aiming to transform physical quantities into basic hidden semantic features. The second-level feature extraction focuses on deep feature deconstruction and spatial dimensionality mapping, using a larger feature dimension to expand and decorrelate the entangled features in the first level, thereby capturing longer-range physical causal laws and providing a high-dimensional vector expression with higher recognizability for the subsequent expert routing unit.

[0070] Through this double-layered and progressively dimensional architecture, the high-dimensional encoding vector not only integrates long-term temporal memory but also possesses extremely high feature discriminative power. This high-dimensional feature representation provides more refined input for subsequent routing units, enabling routing decisions to more accurately identify the specific stage of the current fracturing operation (such as the water diversion stage, sand addition stage, or pump shutdown stage) based on the expanded feature space. This feature evolution process from low to high and from shallow to deep ensures that the model can fully capture extremely subtle trend fluctuations in the fracturing operation data, providing a robust and information-rich feature base for ultimately achieving high-precision pump pressure prediction curves.

[0071] By employing a dual-layer LSTM encoding strategy with progressively increasing dimensionality, and utilizing a spatial mapping mechanism of d_2>d_1, the technical challenge of distinguishing features in low-dimensional space for multi-dimensional fracturing construction data is effectively addressed. This dimensionality-increasing process not only enhances the model's ability to represent nonlinear physical laws but also provides more discriminative input for expert scheduling in the subsequent MoE architecture through feature decoupling. This significantly improves the model's fitting accuracy for complex working conditions, thereby enabling more accurate predictions of pump pressure evolution trends.

[0072] In some embodiments, the feature decoding module using a preset pump pressure prediction model performs correlation feature extraction processing on the high-dimensional encoded vector with causal constraints to determine the correlation feature vector. In specific implementations, the method may further include the following: S1: Using the normalization layer in the feature decoding module, determine the normalized feature vector based on the high-dimensional encoding vector; S2: Using the self-attention layer in the feature decoding module, feature association calculation is performed on the normalized feature vector to determine the time-related feature vector; wherein, the self-attention layer is provided with a time causal mask; S3: Based on the time-related feature vector and the high-dimensional encoding vector, perform summation processing using a preset residual connection structure to determine the related feature vector.

[0073] Specifically, after receiving the high-dimensional encoded vector representing the complex fracturing conditions, the feature decoding module first performs standardization processing by the normalization layer. This process normalizes the mean and scales the variance of each feature dimension in the high-dimensional space, mapping the numerical fluctuation range to a stable distribution interval, thereby generating a normalized feature vector. This processing effectively suppresses gradient vanishing within the deep network, providing a statistically consistent data foundation for subsequent feature correlation calculations across time steps.

[0074] The generated normalized feature vectors are fed into a self-attention layer with a temporal causality mask. During self-attention calculation, this layer uses an upper triangular region filled with a causality mask matrix of negative infinity to filter the attention scores, physically severing the information path between the current moment and future time steps. The temporal correlation feature vectors determined in this way represent each time point only by aggregating the current moment and earlier historical construction states, rigorously simulating the unidirectional temporal evolution logic of the fracturing operation site.

[0075] To further integrate the original signal with deep correlation information, a pre-defined residual connection structure is used to perform element-wise summation between the extracted temporal correlation feature vector and the initial high-dimensional encoded vector. This residual mapping mechanism allows the model to focus solely on learning the residual components in pressure fluctuations while preserving the basic temporal features, thus determining the final correlation feature vector. By constructing this skip-like connection path, the model ensures that it maintains the efficiency of key physical feature transfer and gradient stability even when processing long-range construction sequences.

[0076] By introducing a temporal causal mask into the self-attention layer, the model is forced to follow the temporal causal laws of the fracturing physical process, completely eliminating the information leakage problem during model training. Combined with normalization and residual connection logic, this significantly enhances the model's ability to capture the nonlinear fluctuations in pump pressure, enabling the associated feature vectors to more accurately represent the deep physical logic between construction displacement, sand ratio changes, and pressure response. This provides core feature support for achieving high-precision, unbiased pump pressure curve prediction.

[0077] In some embodiments, the routing unit utilizing a preset pump pressure prediction model determines the allocation weights for the plurality of expert blocks based on the associated feature vectors, and determines a plurality of target expert blocks from the plurality of expert blocks based on the allocation weights. In specific implementations, the method may further include the following: S1: Using the routing unit, perform feature mapping processing on the associated feature vector to determine the initial weight values ​​for each expert block; S2: Using a preset activation function, normalize the initial weight values ​​to determine the assigned weights for each expert block; S3: Filter the numerical relationship of the allocated weights to determine multiple target expert blocks from the multiple expert blocks.

[0078] Specifically, after receiving the associated feature vectors incorporating causal constraint logic, the routing unit performs a linear transformation using an internally deployed weight learning matrix, projecting the high-dimensional feature space onto a score space corresponding to the number of experts. During this process, hidden information such as construction displacement fluctuations and pressure trends contained in the associated feature vectors is mapped into original contribution scores for each expert block. This mapping process is essentially a feature identification of the current fracturing condition, aiming to preliminarily assess the adaptability of each expert subnetwork to the current construction state, and thus determine the initial weight values ​​for each expert block.

[0079] To establish clear competitive or collaborative relationships among multiple experts, a pre-defined Softmax activation function is used to perform non-linear normalization calculations on the initial weight values. This operation transforms the raw scores of each dimension into a probability distribution with values ​​ranging from 0 to 1 and summing to 1, thereby determining the weights assigned to each expert block. Through this probabilistic representation, the model can accurately represent the importance of different expert blocks to the current prediction task in numerical form. For example, during periods of stable construction pressure, the expert responsible for steady-state feature processing will receive a higher weight ratio.

[0080] After obtaining the distribution of all assigned weights, the expert blocks are sorted in descending order of their values, and a Top-k algorithm is used to select the top N experts with the highest weights (e.g., the top two experts by weight). This filtering process sets the weights of expert blocks with lower allocations or weaker correlations to zero or masks them, thus accurately identifying multiple target expert blocks from numerous candidate units. This sparse activation mechanism not only ensures that the model can utilize its most suitable "specialized knowledge" for the current working condition for stress prediction but also significantly reduces the computational overhead of complex deep learning architectures during real-time inference.

[0081] Dynamic weight allocation and target expert selection through routing units enable the model to schedule the most suitable computing resources in real time based on the rapidly changing working conditions at the fracturing site (such as starting pressure and displacement adjustment). This sparse activation mechanism overcomes the limitation of insufficient fitting ability of a single model when faced with complex and multi-dimensional construction data. It ensures that the prediction model can both improve the local accuracy of the predicted pump pressure curve by utilizing the specialized division of labor of expert blocks and maintain high-efficiency computing performance, providing more targeted and real-time technical support for risk early warning at the fracturing site.

[0082] In some embodiments, the routing unit utilizing a preset pump pressure prediction model performs weighted fusion processing on the output of the target expert block according to the assigned weights to determine the decoding feature vector. In specific implementations, the method may further include the following: S1: Based on the associated feature vector, feature transformation processing is performed on each of the target expert blocks to determine the output result of each target expert block; S2: Using the routing unit, the output results of each target expert block and the allocated weights are weighted and fused to determine the hybrid expert feature vector; S3: Based on the hybrid expert feature vector and the associated feature vector, the decoding feature vector is determined using a preset residual connection structure.

[0083] Specifically, during the expert scheduling and feature synthesis phase, multiple selected target expert blocks receive the associated feature vectors output by the feature decoding module and use them as input for parallel computation to perform deep nonlinear feature transformation processing. Each target expert block is internally composed of a multi-layer feedforward neural network (FFN). By performing weight matrix mapping and activation function operations on this feature dimension, the abstract information about construction history and causal constraints in the associated feature vectors is further semantically deconstructed. Since different expert blocks are guided to learn different distributions of working condition samples in fracturing operations (such as stable fracturing conditions, proppant-filled fluctuating conditions, etc.) during the training phase, each expert block will produce expert processing results representing specific physical logic, thereby achieving differentiated capture of the complex nonlinear evolution laws in fracturing operations.

[0084] The routing unit retrieves the assigned weights determined in the preceding steps and performs dynamic aggregation on the expert processing results produced by each target expert block. By performing a dot product operation on each expert processing result and its corresponding assigned weight, a series of weighted feature components are obtained. These components are then accumulated in their corresponding spatial dimensions to determine a hybrid expert feature vector representing the global feature estimation of the current construction environment. This fusion process demonstrates the routing unit's ability to dynamically intervene in the calculation path, that is, to adjust the contribution ratio of each expert's "suggestion" in real time according to the fluctuation characteristics of the current displacement and pressure, ensuring that the prediction model can make decisions based on the expert combination that best fits the current actual working conditions.

[0085] To preserve the stability of the original features during expert knowledge reorganization, a pre-defined residual connection structure is used to perform element-wise summation between the generated hybrid expert feature vector and the original associated feature vector before entering the expert hybrid module. This skip-connection mechanism allows the model to superimpose the operating condition correction information learned by the expert blocks with the basic causal association features, thereby determining the final decoded feature vector. This approach not only effectively prevents gradient vanishing or feature degradation problems that may occur when deep neural networks process long-term fracturing data, but also achieves a progressively refined expression of pump pressure prediction features through residual mapping, providing a robust feature base for the final output of a high-confidence pump pressure prediction curve.

[0086] By employing a routing unit-driven expert weighted fusion and residual connection mechanism, targeted modeling of pump pressure evolution under various fracturing conditions was achieved, effectively addressing the problem of insufficient fitting ability of a single linear prediction model when faced with drastic changes in construction parameters. This dynamically weighted hybrid expert architecture not only enhances the semantic representation depth of the model for multi-dimensional data from the construction site but also utilizes the residual structure to ensure the integrity of physical causal information during feature fusion, thereby significantly improving the accuracy and robustness of predicted pump pressure curves in the stimulation of complex shale oil and gas reservoirs.

[0087] In some embodiments, the preset pump pressure prediction model further includes a linear layer. The method, which determines the pump pressure prediction result within a target time period based on the decoded feature vector through dimensional mapping processing, may further include the following in its specific implementation: S1: Using the linear layer in the preset pump pressure prediction model, perform the dimension mapping process on the decoded feature vector to determine the predicted feature vector; S2: Determine the target time period relative to the current time based on the first preset duration corresponding to the predicted feature vector; S3: Map the numerical points contained in the predicted feature vector to each sampling time within the target time period to determine the pump pressure prediction result within the target time period.

[0088] Specifically, the decoded feature vectors aggregate multi-dimensional operating conditions and causal relationships, but their form remains in a high-dimensional abstract feature space. Using a linear layer deployed in the pre-defined pump pressure prediction model, a projection transformation from the high-dimensional hidden space to the low-dimensional physical numerical space is performed. Through weight matrix operations, the deep features contained in the decoded feature vectors are linearly combined and their dimensions compressed to determine the predicted feature vectors. This process achieves a cross-dimensional transformation of data from "model logical features" to "physical evolution trends," ensuring that the output results are strictly aligned mathematically with the expected number of pump pressure sampling points.

[0089] The number of elements in the generated predicted feature vector matches the model's preset prediction horizon (i.e., the first preset duration). If the first preset duration is set to 300 seconds into the future, the predicted feature vector contains a corresponding number of regression values. By using the current moment as the starting anchor point, this 300-second time window is extended into the future time domain after the current moment, thereby precisely locking the target time period relative to the current moment. This step assigns a clear time coordinate to the predicted values, establishing the specific position of the predicted pressure curve on the fracturing operation time axis.

[0090] After defining the time limit, point-to-point mapping is performed using the numerical points arranged in the predicted feature vector and the sampling moments distributed at a fixed frequency within the target time period. By associating the i-th component in the vector with the i-th sampling instant within the target time period, the discrete regression values ​​are transformed into physically meaningful, time-varying pump pressure prediction results. This mapping process reconstructs the dynamic fluctuation trajectory of pump pressure in future periods, directly outputting a numerical sequence that can be used for pressure early warning and risk decision-making at the fracturing command site.

[0091] The dimensional mapping and time anchoring mechanism implemented through linear layers achieves a precise conversion from abstract high-dimensional features to specific physical values, ensuring that the pump pressure prediction results have both deep logical support and clear spatiotemporal attributes. Utilizing point-to-point mapping processing, the model can output high-resolution future pump pressure evolution curves, effectively solving the problems of lag and insufficient prediction resolution in traditional physical simulation methods during complex shale formation construction. This provides a quantitative and intuitive technical basis for real-time risk management at fracturing sites.

[0092] In some embodiments, within the model, after the temporal memory is extracted by LSTM and the data is deeply processed by the MoE expert block, the data exists in a high-dimensional hidden feature space. At this time, the "decoded feature vector" is a series of highly abstract numerical values ​​(e.g., 512-dimensional or 1024-dimensional), which represent the model's deep understanding of the current working conditions, historical trends, and causal logic, but these numerical values ​​themselves do not have intuitive physical meaning.

[0093] Engineers at fracturing sites need specific future pump pressure values ​​with physical units (such as MPa). Dimensional mapping transforms the complex mathematical logic of the model into measurable pressure readings in the real world, thus bridging the dimensional gap between abstract features and concrete physical quantities.

[0094] The core purpose of dimensionality mapping (usually implemented through a linear layer) is to perform a precise spatial projection. First, it compresses or reassembles the high-dimensional hidden vector into a low-dimensional vector consistent with the prediction time step through weight matrix operations, ensuring that the output is strictly aligned in number with the preset sampling points for the "future 300 seconds" or "future 600 seconds".

[0095] Secondly, the mapping process is essentially a regression transformation from nonlinear features to linear values. It utilizes trained weight parameters to perform a final weighted combination of the features aggregated from each expert block, removing redundant computational noise and extracting the signal components that best reflect the trend of pressure evolution. This purposeful dimensionality transformation ensures that every numerical point in the model's output accurately corresponds to the physical pressure at a specific future moment.

[0096] Dimensional mapping processing achieves the following significant benefits: Significantly improved prediction accuracy: Because the mapping process involves secondary feature selection and numerical fitting based on fully decoded features, it can capture subtle pressure fluctuation precursors in high-dimensional features. The output pump pressure prediction curve closely matches the actual physical response of the formation compared to traditional simple linear extrapolation. Achieving high-resolution real-time early warning: Mapping processing directly produces continuous time series data, rather than single numerical points. This high-resolution output allows field technicians to clearly see the pressure evolution trend over a future period (such as whether a sudden surge in sand blockage pressure will occur), thus providing sufficient reaction time for risk prevention and control.

[0097] In some embodiments, the method may further include the following: S1: Based on the pump pressure prediction results, determine the abnormal construction conditions at the fracturing construction site; S2: Based on the abnormal construction status, determine the corresponding risk warning information; S3: Based on the risk warning information, determine the adjustment strategy for the construction plan at the fracturing construction site.

[0098] Specifically, the predicted pump pressure for the target time period is input into a preset safety criterion. By analyzing the curvature changes and extreme value distribution of the prediction curve, it is determined whether the fracturing operation site is in a potentially abnormal state. Specifically, when the prediction curve shows that the pump pressure will continue to rise over a future period and the rate of increase exceeds a preset slope threshold, a potential sand blockage risk is identified; if the predicted pressure value exceeds the safe rated pressure of the wellhead equipment, it is determined to be an overpressure warning state. This process achieves a logical leap from a purely numerical sequence to a construction safety evaluation with physical meaning.

[0099] For specific construction anomalies identified, a pre-set risk correlation database is searched to match corresponding risk warning information. This warning information not only includes the specific category of the anomaly (such as ground compaction, wellbore obstruction, or equipment failure), but also covers the risk trigger countdown calculated based on the prediction timeline and the expected pressure peak. This approach visualizes the high-dimensional output of the prediction model into warning instructions with time window indications, providing on-site workers with clear psychological expectations and a buffer period for response.

[0100] Based on the severity and urgency of the risk warning information, targeted adjustment strategies for the construction plan are retrieved and determined from a pre-set engineering expert knowledge base. When facing predicted high-risk sand blockage conditions, the generated strategies can directly intervene in specific pumping parameters, such as reducing the construction discharge rate according to preset steps, reducing the real-time addition concentration of proppant, or increasing the proportion of slickwater for pressure-limited circulation.

[0101] By constructing a closed-loop logic from pump pressure prediction results to construction plan adjustments, the problems of lagging human experience judgment and strong subjectivity in risk management in existing technologies are effectively solved. Utilizing forward-looking data support provided by the prediction model, early detection and quantitative warning of abnormal construction conditions are achieved. This allows for proactive intervention in displacement and sand ratio based on precisely generated adjustment strategies, significantly reducing the risks of sand blockage and overpressure during fracturing operations, and ensuring the continuity of fracturing operations and well control safety.

[0102] As can be seen from the above, the pump pressure prediction method provided in this specification involves acquiring construction data from a target fracturing construction site, determining an initial feature sequence based on the construction data, determining distribution feature parameters within a corresponding preset sliding window based on the initial feature sequence, and normalizing the initial feature sequence using the distribution feature parameters to determine a target feature sequence. The target feature sequence is then encoded and converted using a feature encoding module of a preset pump pressure prediction model to determine a high-dimensional encoding vector. The preset pump pressure prediction model includes at least a feature encoding module, a feature decoding module, a routing unit, and multiple expert blocks. The feature decoding module of the preset pump pressure prediction model is used to normalize the initial feature sequence to determine a target feature sequence. The high-dimensional encoded vector is subjected to causal constraint-based correlation feature extraction processing to determine the correlation feature vector; using the routing unit of the preset pump pressure prediction model, the weights assigned to the multiple expert blocks are determined according to the correlation feature vector, and multiple target expert blocks are determined from the multiple expert blocks according to the weights; using the routing unit of the preset pump pressure prediction model, the output results of the target expert blocks are weighted and fused according to the weights to determine the decoded feature vector; wherein, the output result of the target expert block is the result obtained by processing the correlation feature vector using the target expert block; based on the decoded feature vector, the pump pressure prediction result within the target time period is determined through dimension mapping processing. In this way, by acquiring construction data from the target fracturing construction site and performing normalization processing using distribution feature parameters, the subjectivity problem caused by reliance on human experience in existing technologies is overcome. By utilizing the feature encoding module, the feature decoding module with causal constraint-based correlation feature extraction processing, and the routing unit's weighted fusion processing of the output results of multiple target expert blocks, in-depth mining and dynamic scheduling of high-dimensional encoded vectors and correlation feature vectors in multi-dimensional construction data from the site are achieved. This solves the problem that physical simulation methods cannot make real-time predictions for multi-dimensional data, thereby achieving more accurate pump pressure prediction.

[0103] See Figure 2 As shown in the embodiments of this specification, a specific electronic device is also provided, wherein the electronic device includes a network communication port 201, a processor 202 and a memory 203, and the above structures are connected by internal cables so that the various structures can perform specific data interaction.

[0104] Specifically, the network communication port 201 can be used to acquire construction data at the target fracturing construction site and determine an initial feature sequence based on the construction data.

[0105] The processor 202 is specifically configured to: determine distribution feature parameters within a preset sliding window based on the initial feature sequence; normalize the initial feature sequence using the distribution feature parameters to determine a target feature sequence; and perform encoding transformation on the target feature sequence using a feature encoding module of a preset pump pressure prediction model to determine a high-dimensional encoding vector. The preset pump pressure prediction model includes at least a feature encoding module, a feature decoding module, a routing unit, and multiple expert blocks. The feature decoding module of the preset pump pressure prediction model is used to extract causal-constrained correlation features from the high-dimensional encoding vector. The process involves: processing and determining the associated feature vector; using the routing unit of a preset pump pressure prediction model, determining the allocation weights for the multiple expert blocks based on the associated feature vectors, and identifying multiple target expert blocks from the multiple expert blocks based on the allocation weights; using the routing unit of the preset pump pressure prediction model, performing weighted fusion processing on the output results of the target expert blocks based on the allocation weights to determine the decoded feature vectors; wherein, the output results of the target expert blocks are the results obtained by processing the associated feature vectors using the target expert blocks; and based on the decoded feature vectors, determining the pump pressure prediction results within the target time period through dimension mapping processing.

[0106] The memory 203 can be used to store the corresponding instruction program.

[0107] Based on the above method, the relevant structural performance of electronic devices can be effectively utilized to improve the data processing speed of electronic devices and efficiently realize a pump pressure prediction method.

[0108] In this embodiment, the network communication port 201 can be a virtual port bound to different communication protocols, thereby enabling the sending or receiving of different data. For example, the network communication port can be a port responsible for web data communication, a port responsible for FTP data communication, or a port responsible for email data communication. Furthermore, the network communication port can also be a physical communication interface or communication chip. For example, it can be a wireless mobile network communication chip, such as GSM or CDMA; it can also be a Wi-Fi chip; or it can be a Bluetooth chip.

[0109] In this embodiment, the processor 202 can be implemented in any suitable manner. For example, the processor can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers, etc. This specification is not limiting.

[0110] In this embodiment, the memory 203 may include a hierarchy. In a digital system, anything that can store binary data can be a memory. In an integrated circuit, a circuit with storage function but no physical form is also called a memory, such as RAM, FIFO, etc. In a system, a storage device with a physical form is also called a memory, such as a memory stick, TF card, etc.

[0111] This specification also provides a computer-readable storage medium based on the above-described pump pressure prediction method, which acquires construction data from a target fracturing construction site and determines an initial feature sequence based on the construction data; determines distribution feature parameters within a corresponding preset sliding window based on the initial feature sequence, and normalizes the initial feature sequence using the distribution feature parameters to determine a target feature sequence; and uses a feature encoding module of a preset pump pressure prediction model to encode and transform the target feature sequence to determine a high-dimensional encoding vector; wherein the preset pump pressure prediction model includes at least: a feature encoding module, a feature decoding module, a routing unit, and multiple expert blocks; and utilizes the feature decoding module of the preset pump pressure prediction model... The module performs causal constraint-based correlation feature extraction processing on the high-dimensional encoded vector to determine the correlation feature vector; using the routing unit of the preset pump pressure prediction model, it determines the allocation weights for the multiple expert blocks based on the correlation feature vectors, and determines multiple target expert blocks from the multiple expert blocks based on the allocation weights; using the routing unit of the preset pump pressure prediction model, it performs weighted fusion processing on the output results of the target expert blocks based on the allocation weights to determine the decoded feature vector; wherein, the output result of the target expert block is the result obtained by processing the correlation feature vector using the target expert block; based on the decoded feature vector, the pump pressure prediction result within the target time period is determined through dimension mapping processing.

[0112] In this embodiment, the storage medium includes, but is not limited to, Random Access Memory (RAM), Read-Only Memory (ROM), Cache, Hard Disk Drive (HDD), or Memory Card. The memory can be used to store computer program instructions. The network communication unit can be an interface configured according to standards specified in the communication protocol for network connection communication.

[0113] In this embodiment, the specific functions and effects implemented by the program instructions stored in the computer-readable storage medium can be explained in comparison with other embodiments, and will not be repeated here.

[0114] See Figure 3At the software level, this specification also provides a pump pressure prediction device, which may specifically include the following structural modules: The initial sequence determination module 301 is used to acquire construction data at the target fracturing construction site and determine an initial feature sequence based on the construction data. The target sequence determination module 302 is used to determine the distribution feature parameters within a corresponding preset sliding window based on the initial feature sequence, and to normalize the initial feature sequence using the distribution feature parameters to determine the target feature sequence. The first vector determination module 303 is used to encode and transform the target feature sequence using the feature encoding module of the preset pump pressure prediction model to determine the high-dimensional encoding vector; wherein, the preset pump pressure prediction model includes at least: a feature encoding module, a feature decoding module, a routing unit, and multiple expert blocks; The second vector determination module 304 is used to perform correlation feature extraction processing with causal constraints on the high-dimensional encoded vector using the feature decoding module of the preset pump pressure prediction model to determine the correlation feature vector. The expert block determination module 305 is used to determine the allocation weights for the multiple expert blocks based on the associated feature vectors using the routing unit of the preset pump pressure prediction model, and to determine multiple target expert blocks from the multiple expert blocks based on the allocation weights. The third vector determination module 306 is used to utilize the routing unit of the preset pump pressure prediction model to perform weighted fusion processing on the output result of the target expert block according to the allocated weights, and determine the decoded feature vector; wherein, the output result of the target expert block is the result obtained by processing the associated feature vector using the target expert block; The result determination module 307 is used to determine the pump pressure prediction result within the target time period based on the decoded feature vector through dimension mapping processing.

[0115] In some embodiments, the target sequence determination module 302, in its specific implementation, uses a preset sliding window to divide the initial feature sequence into multiple sub-feature sequences; wherein, one sub-feature sequence corresponds to one time step; based on each sub-feature sequence, the average value and standard deviation of the features at each time step are calculated as distribution feature parameters for each time step; using the distribution feature parameters for each time step, the sub-feature sequences at the corresponding time steps are normalized to obtain the target feature sequence.

[0116] In some embodiments, the first vector determination module 303, in its specific implementation, utilizes the first layer of long short-term memory neural network in the feature encoding module to perform first-level feature extraction on the target feature sequence to determine a first encoded feature vector with a first preset dimension; and utilizes the second layer of long short-term memory neural network in the feature encoding module to perform second-level feature extraction on the first encoded feature vector to determine the high-dimensional encoded vector with a second preset dimension; wherein the second preset dimension is greater than the first preset dimension.

[0117] In some embodiments, the second vector determination module 304, in its specific implementation, utilizes the normalization layer in the feature decoding module to determine a normalized feature vector based on the high-dimensional encoded vector; utilizes the self-attention layer in the feature decoding module to perform feature association calculation on the normalized feature vector to determine a time-related feature vector; wherein, the self-attention layer is provided with a time causal mask; and based on the time-related feature vector and the high-dimensional encoded vector, performs summation processing using a preset residual connection structure to determine the related feature vector.

[0118] In some embodiments, the expert block determination module 305, in its specific implementation, utilizes the routing unit to perform feature mapping processing on the associated feature vector to determine the initial weight value for each expert block; utilizes a preset activation function to normalize the initial weight value to determine the allocated weight for each expert block; and performs filtering processing on the numerical relationship of the allocated weight to determine multiple target expert blocks from the multiple expert blocks.

[0119] In some embodiments, the third vector determination module 306, in its specific implementation, performs feature transformation processing on each of the target expert blocks according to the associated feature vector to determine the output result of each target expert block; uses the routing unit to perform weighted fusion processing on the output result of each target expert block and the allocated weight to determine the hybrid expert feature vector; and uses a preset residual connection structure to determine the decoding feature vector based on the hybrid expert feature vector and the associated feature vector.

[0120] In some embodiments, the result determination module 307, in specific implementation, uses the linear layer in the preset pump pressure prediction model to perform the dimension mapping process on the decoded feature vector to determine the predicted feature vector; determines the target time period relative to the current time based on the first preset duration corresponding to the predicted feature vector; and performs mapping processing between the numerical points contained in the predicted feature vector and each sampling time within the target time period to determine the pump pressure prediction result within the target time period.

[0121] In some embodiments, the above-described device further includes: determining an abnormal construction state at the fracturing construction site based on the pump pressure prediction result; determining corresponding risk warning information based on the abnormal construction state; and determining an adjustment strategy for the construction plan at the fracturing construction site based on the risk warning information.

[0122] It should be noted that the units, devices, or modules described in the above embodiments can be implemented by computer chips or physical entities, or by products with certain functions. For ease of description, the above devices are described by dividing them into various modules according to their functions. Of course, in implementing this specification, the functions of each module can be implemented in the same software and / or hardware, or modules that implement the same function can be implemented by a combination of sub-modules or sub-units, etc. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and there may be other division methods in actual implementation. For example, units or components can be combined or integrated into another system, or some features can be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection between the devices or units shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0123] As can be seen from the above, based on the pump pressure prediction device provided in the embodiments of this specification, construction data of the target fracturing construction site is acquired, and an initial feature sequence is determined according to the construction data; based on the initial feature sequence, distribution feature parameters within a corresponding preset sliding window are determined, and the initial feature sequence is normalized using the distribution feature parameters to determine the target feature sequence; the target feature sequence is encoded and converted using the feature encoding module of a preset pump pressure prediction model to determine a high-dimensional encoding vector; wherein, the preset pump pressure prediction model includes at least: a feature encoding module, a feature decoding module, a routing unit, and multiple expert blocks; the feature decoding module of the preset pump pressure prediction model is used to... The high-dimensional encoded vector undergoes causal constraint-based correlation feature extraction processing to determine the correlation feature vector. Using the routing unit of a preset pump pressure prediction model, weights are assigned to the multiple expert blocks based on the correlation feature vector, and multiple target expert blocks are identified from the multiple expert blocks based on the assigned weights. The routing unit of the preset pump pressure prediction model performs weighted fusion processing on the output results of the target expert blocks based on the assigned weights to determine the decoded feature vector. The output result of the target expert block is the result obtained by processing the correlation feature vector using the target expert block. Based on the decoded feature vector, the pump pressure prediction result within the target time period is determined through dimensionality mapping processing.

[0124] In a specific scenario example, the pump pressure prediction method and device provided in this specification can be applied to solve the technical problems in the prior art where pump pressure prediction relies too heavily on human experience, leading to high subjectivity, and physical simulation methods have a lag, making it impossible to make real-time and accurate predictions based on multi-dimensional construction data on site. The specific implementation process may include the following:

[0125] At a shale oil fracturing construction site, real-time sensor data from the sand mixer and pump truck were collected, acquiring data including 12m... 3 The system measures multiple physical quantities, including the construction displacement per minute and the construction hydraulic pressure at 80 MPa. Through time axis alignment, the system synchronizes the displacement and pressure signals with varying sampling frequencies and dynamically calculates the average and standard deviation of the current operating conditions using a preset 600-second sliding window. These distributional characteristic parameters are then used to normalize the initial feature sequence, transforming fluctuating physical values ​​into a standardized target feature sequence, providing a clean data foundation for subsequent deep inference of large models.

[0126] After the target feature sequence enters the LSTM-MoE fusion model, the feature encoding module performs spatial dimensionality upscaling mapping through a two-layer LSTM structure, transforming the data into a 512-dimensional high-dimensional encoding vector to capture long-range temporal dependencies in the fracturing process. In the feature decoding stage, a self-attention layer with causal masking strictly shields future information, ensuring that the extraction of associated features conforms to the logical sequence of physical time. Subsequently, the routing unit dynamically allocates target expert blocks adept at handling this condition for weighted fusion based on the current feature's tendency towards "high sand ratio fluctuations." Finally, through dimensionality mapping in a linear layer, a predicted pump pressure curve reflecting the trend over the next 300 seconds is output.

[0127] When the predicted pump pressure curve indicates that the pressure will continue to rise and exceed the safety threshold of 95 MPa within the next two minutes, the system immediately identifies the abnormal construction state of "potential sand blockage". In response to this anomaly, the system automatically triggers a high-risk warning and pushes an adjustment strategy for the construction plan to the central control interface, recommending that on-site technicians reduce the proppant concentration by 20% and increase the pumping rate in a stepwise manner. This proactive warning and intervention mechanism allows the construction team to shift from traditional pump shutdown after pressure exceeds the limit to trend-based pre-emptive control, effectively avoiding downhole sand blockage risks and ensuring the continuity and safety of reservoir stimulation operations.

[0128] In some embodiments, see Figure 4The diagram illustrates the overall hierarchical architecture of a pre-defined pump pressure prediction model provided in this embodiment. The model employs an encoder-decoder topology, specifically consisting of an upper-layer feature encoding module composed of a two-layer long short-term memory neural network (LSTM), and a lower-layer feature decoding module composed of multiple stacked decode blocks. This deep heterogeneous architecture enables the layer-by-layer deconstruction and semantic reshaping of the nonlinear physical features in fracturing construction data.

[0129] Specifically, the feature encoding module receives the preprocessed target feature sequence. See also... Figure 4 The hierarchical correspondence in the model involves the target feature sequence first being input into a first-layer Long Short-Term Memory (LSTM) neural network for primary temporal feature capture. The resulting intermediate feature vector is then directly used as the input to a second-layer LSM. In this two-layer cascaded process, the number of hidden nodes in the second-layer neural network is adjusted to achieve a dimensionality upgrade from the low-dimensional physical space to the high-dimensional feature space. The high-dimensional encoding vector output by the second-layer LSM, with a preset dimension (e.g., 512 dimensions), not only contains long-distance dependencies of fracturing operation parameters over time but also provides a rich information carrier for subsequent fine-grained feature decoding through dimensional expansion.

[0130] The feature decoding module consists of multiple vertically stacked decoding blocks. Figure 4 The progressive processing logic is illustrated in a series. The high-dimensional encoded vector produced by the feature encoding module serves as the initial input to the decoding module, flowing sequentially through each level of the decoding block. Each decoding block integrates normalization, a self-attention mechanism with causal masking, and hybrid expert (MoE) processing logic (its internal structure is not detailed in the figure). Through the iterative refinement and residual correction of associated features by each level of the decoding block, the abstract high-dimensional encoded vector is gradually transformed into a decoded feature vector with physical evolution directionality. This multi-block stacked architecture ensures that the model has sufficient nonlinear fitting depth, effectively capturing the drastic pump pressure fluctuations caused by formation heterogeneity during fracturing operations.

[0131] At the very end of the model, the decoded feature vector output by the final decoding block is passed to the subsequent dimension mapping layer. Through linear mapping, the decoded deep features are projected onto the pressure numerical space of the target time period, ultimately determining the pump pressure prediction result. This fusion-based large model architecture, constructed collaboratively by a two-layer LSTM and multi-level decoding blocks, greatly enhances the model's semantic representation ability of complex multi-dimensional data from construction sites while ensuring temporal causality, laying a solid algorithmic foundation for achieving more accurate real-time pump pressure prediction.

[0132] While this specification provides the steps of operation for the methods described in the embodiments or flowcharts, more or fewer steps may be included based on conventional or non-inventive means. The order of steps listed in the embodiments is merely one possible order of execution among many steps and does not represent the only possible order. In actual device or client product execution, the methods shown in the embodiments or drawings may be executed sequentially or in parallel (e.g., in a parallel processor or multi-threaded processing environment, or even a distributed data processing environment). The terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, product, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, product, or apparatus. Without further limitations, the presence of other identical or equivalent elements in a process, method, product, or apparatus that includes said elements is not excluded. The terms "first," "second," etc., are used to denote names and do not indicate any particular order.

[0133] Those skilled in the art will also know that, besides implementing the controller using purely computer-readable program code, the same functions can be achieved by logically programming the method steps, making the controller function as logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers (PLCs), and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the devices within it used to implement various functions can also be considered structures within that hardware component. Alternatively, the devices used to implement various functions can be considered as both software modules implementing the method and structures within a hardware component.

[0134] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this specification can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions of this specification can essentially be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, mobile terminal, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments of this specification.

[0135] Although this specification has been described by way of examples, those skilled in the art will recognize that many variations and modifications are possible without departing from the spirit of this specification, and it is intended that the appended claims cover such variations and modifications without departing from the spirit of this specification.

Claims

1. A method for predicting pump pressure, characterized in that, include: Acquire construction data from the target fracturing site, and determine an initial feature sequence based on the construction data; Based on the initial feature sequence, the distribution feature parameters within the corresponding preset sliding window are determined, and the initial feature sequence is normalized using the distribution feature parameters to determine the target feature sequence. The target feature sequence is encoded and converted using the feature encoding module of the preset pump pressure prediction model to determine a high-dimensional encoding vector; wherein the preset pump pressure prediction model includes at least: a feature encoding module, a feature decoding module, a routing unit, and multiple expert blocks; Using the feature decoding module of the preset pump pressure prediction model, the high-dimensional encoded vector is subjected to correlation feature extraction processing with causal constraints to determine the correlation feature vector; Using the routing unit of the preset pump pressure prediction model, the allocation weights for the multiple expert blocks are determined according to the associated feature vectors, and multiple target expert blocks are determined from the multiple expert blocks according to the allocation weights. The routing unit of the preset pump pressure prediction model performs weighted fusion processing on the output of the target expert block according to the allocated weights to determine the decoding feature vector; wherein, the output of the target expert block is the result obtained by processing the associated feature vector using the target expert block; Based on the decoded feature vector, the pump pressure prediction result within the target time period is determined through dimensional mapping processing.

2. The method according to claim 1, characterized in that, The step of determining the distribution feature parameters within a corresponding preset sliding window based on the initial feature sequence, and then normalizing the initial feature sequence using the distribution feature parameters to determine the target feature sequence, includes: Using a preset sliding window, the initial feature sequence is divided into multiple sub-feature sequences; where each sub-feature sequence corresponds to one time step. Based on each sub-feature sequence, the mean and standard deviation of the features at each time step are calculated and used as the distribution feature parameters at each time step. By utilizing the distribution characteristic parameters of each time step, the sub-feature sequences of the corresponding time steps are normalized to obtain the target feature sequence.

3. The method according to claim 2, characterized in that, The preset pump pressure prediction model is a fusion model of LSTM-MoE. The feature encoding module of the preset pump pressure prediction model performs encoding transformation processing on the target feature sequence to determine a high-dimensional encoding vector, including: Using the first layer of the long short-term memory neural network in the feature encoding module, the first-level feature extraction is performed on the target feature sequence to determine the first encoded feature vector with a first preset dimension; Using the second layer of long short-term memory neural network in the feature encoding module, a second-level feature extraction is performed on the first encoded feature vector to determine the high-dimensional encoded vector with a second preset dimension; Wherein, the second preset dimension is greater than the first preset dimension.

4. The method according to claim 3, characterized in that, The feature decoding module, utilizing a pre-defined pump pressure prediction model, performs causal constraint-based correlation feature extraction processing on the high-dimensional encoded vector to determine the correlation feature vector, including: Using the normalization layer in the feature decoding module, a normalized feature vector is determined based on the high-dimensional encoded vector; The self-attention layer in the feature decoding module is used to perform feature association calculation on the normalized feature vector to determine the temporal correlation feature vector; wherein, the self-attention layer is provided with a temporal causal mask; The associated feature vector is determined by summing the time-related feature vector and the high-dimensional encoding vector using a preset residual connection structure.

5. The method according to claim 4, characterized in that, The routing unit utilizing the preset pump pressure prediction model determines the allocation weights for the multiple expert blocks based on the associated feature vectors, and determines multiple target expert blocks from the multiple expert blocks based on the allocation weights, including: Using the routing unit, feature mapping processing is performed on the associated feature vector to determine the initial weight values ​​for each expert block; The initial weight values ​​are normalized using a preset activation function to determine the assigned weights for each expert block. The numerical relationships of the assigned weights are filtered to determine multiple target expert blocks from the multiple expert blocks.

6. The method according to claim 5, characterized in that, The routing unit utilizing the preset pump pressure prediction model performs weighted fusion processing on the output of the target expert block according to the assigned weights to determine the decoding feature vector, including: Based on the associated feature vector, feature transformation processing is performed on each of the target expert blocks to determine the output result of each target expert block; Using the routing unit, the output results of each target expert block and the assigned weights are weighted and fused to determine the hybrid expert feature vector; Based on the hybrid expert feature vector and the associated feature vector, the decoding feature vector is determined using a preset residual connection structure.

7. The method according to claim 6, characterized in that, The preset pump pressure prediction model further includes a linear layer. The step of determining the pump pressure prediction result within the target time period based on the decoded feature vector through dimensionality mapping processing includes: Using the linear layer in the preset pump pressure prediction model, the dimension mapping process is performed on the decoded feature vector to determine the predicted feature vector; Based on the first preset duration corresponding to the predicted feature vector, the target time period relative to the current time is determined; The pump pressure prediction result within the target time period is determined by mapping the numerical points contained in the predicted feature vector to each sampling time within the target time period.

8. The method according to claim 7, characterized in that, The method further includes: Based on the pump pressure prediction results, determine the abnormal construction conditions at the fracturing construction site; Based on the aforementioned abnormal construction conditions, determine the corresponding risk warning information; Based on the risk warning information, a strategy for adjusting the construction plan at the fracturing site is determined.

9. A pump pressure prediction device, characterized in that, include: The initial sequence determination module is used to acquire construction data at the target fracturing construction site and determine the initial feature sequence based on the construction data. The target sequence determination module is used to determine the distribution feature parameters within a corresponding preset sliding window based on the initial feature sequence, and to normalize the initial feature sequence using the distribution feature parameters to determine the target feature sequence. The first vector determination module is used to encode and transform the target feature sequence using the feature encoding module of the preset pump pressure prediction model to determine the high-dimensional encoded vector; wherein, the preset pump pressure prediction model includes at least: a feature encoding module, a feature decoding module, a routing unit, and multiple expert blocks; The second vector determination module is used to perform correlation feature extraction processing with causal constraints on the high-dimensional encoded vector by using the feature decoding module of the preset pump pressure prediction model to determine the correlation feature vector. The expert block determination module is used to determine the allocation weights for the multiple expert blocks based on the associated feature vectors using the routing unit of the preset pump pressure prediction model, and to determine multiple target expert blocks from the multiple expert blocks based on the allocation weights. The third vector determination module is used to utilize the routing unit of the preset pump pressure prediction model to perform weighted fusion processing on the output result of the target expert block according to the allocated weights, and determine the decoded feature vector; wherein, the output result of the target expert block is the result obtained by processing the associated feature vector using the target expert block; The result determination module is used to determine the pump pressure prediction result within the target time period based on the decoded feature vector through dimension mapping processing.

10. An electronic device, characterized in that, It includes a processor and a memory for storing processor-executable instructions, wherein the processor, when executing the instructions, implements the steps of the method according to any one of claims 1 to 8.

Images