A method and system for real-time prediction of flow conditions in an LNG pipeline transportation system

Abstract

This invention relates to the fields of fluid mechanics and pipeline transportation technology, specifically to a method and system for real-time prediction of flow state in an LNG pipeline transportation system. It includes: a multi-source heterogeneous data acquisition and preprocessing module for acquiring and preprocessing raw data from distributed sensors to generate a regular multi-dimensional time-series data stream; a physical mechanism constraint embedding module for constructing the LNG multiphase flow control equations and discretizing them into a physical constraint loss function that can be embedded in a neural network for training; a deep spatiotemporal fusion prediction model module for receiving the data stream and training and inferring based on the physical constraint loss function, outputting prediction results for the velocity field, phase distribution, and pressure fluctuations of the entire pipeline within a future time window; and a real-time prediction and visualization output module for inverse standardization, visualization rendering, and intelligent alarm processing of the prediction results. This invention significantly improves the comprehensiveness of state perception, the generalization ability of the model, and the safety of the system.

Description

Technical Field

[0001] This invention relates to the fields of fluid mechanics and pipeline transportation technology, specifically to a method and system for real-time prediction of flow state in an LNG pipeline transportation system. Background Technology

[0002] Liquefied natural gas (LNG) plays an increasingly important role as a clean energy source in the global energy structure transformation, and its long-distance transportation mainly relies on complex cryogenic pipeline systems. LNG pipeline transportation systems operate under harsh conditions, involving complex physical processes such as multiphase flow, strong unsteady states, and cryogenic environments. The safety and stability of the system are highly dependent on the accurate perception and real-time monitoring of the internal flow state. Currently, monitoring and management technologies in the LNG transportation field still face many challenges, and there is an urgent need for an intelligent method that can efficiently and accurately predict flow states to support safe dispatching and emergency response.

[0003] Among them, the real-time flow state prediction technology for LNG pipeline transportation systems aims to dynamically reconstruct key parameters such as the velocity field, phase distribution, and pressure fluctuations of the fluid within the pipeline by integrating multi-source sensor data and physical models. This technology is the core foundation for realizing intelligent pipeline operation and maintenance, leak early warning, and energy efficiency optimization. Its prediction accuracy and timeliness directly determine the reliability and economy of the entire transportation system.

[0004] Existing technologies for predicting LNG pipeline flow states have significant shortcomings: First, traditional monitoring methods rely heavily on pressure and temperature sensors at discrete points, making it difficult to comprehensively capture the dynamic flow characteristics of the entire pipeline segment. Second, existing models often employ simplified one-dimensional steady-state assumptions, failing to accurately describe the nonlinear evolution of multiphase flow under transient conditions. Third, while data-driven methods have been applied in some scenarios, they generally lack deep integration with physical mechanisms, resulting in weak generalization and poor extrapolation performance. Furthermore, existing physical information neural network methods primarily target simple fluids or fixed boundary conditions, failing to fully consider the latent heat of phase change, the strong nonlinear effects of compressibility and pipe wall friction coupling in low-temperature, long-distance LNG transportation, and the spatiotemporal correlation characteristics under the one-dimensional pipeline topology, leading to poor application results in LNG pipeline scenarios. Finally, existing systems generally lack real-time response capabilities to extreme conditions (such as rapid start-up and shutdown, throttling effects, or external disturbances), making it difficult to meet the high safety requirements of industrial applications. The aforementioned problems severely restrict the development of LNG pipeline transportation systems towards intelligence and autonomy. Therefore, there is an urgent need for a flow state prediction method and system that integrates multi-source heterogeneous data, embeds physical constraints, and has strong real-time prediction capabilities.

[0005] Therefore, existing technologies still need further development. Summary of the Invention

[0006] The purpose of this invention is to overcome the above-mentioned technical deficiencies and provide a method and system for real-time prediction of the flow status of an LNG pipeline transportation system, so as to solve the problems existing in the prior art.

[0007] To achieve the above-mentioned technical objectives, according to a first aspect of the present invention, the present invention provides a real-time flow state prediction system for an LNG pipeline transportation system, comprising: The multi-source heterogeneous data acquisition and preprocessing module is used to acquire raw sensor data from a distributed sensor network and perform cleaning, spatiotemporal alignment and standardization to generate a multi-dimensional time-series data stream. The physical mechanism constraint embedding module is used to construct a set of control equations describing the multiphase flow process in LNG pipelines and discretize them into physical constraint loss functions that can be embedded in neural network training. The deep spatiotemporal fusion prediction model module is used to receive the multidimensional time series data stream, and perform training and inference based on the physical constraint loss function to generate prediction results of the velocity field, phase distribution and pressure fluctuation of the entire pipeline within the future time window; The real-time prediction and visualization output module is used to perform inverse normalization and visualization rendering on the prediction results and generate alarm information.

[0008] Specifically, the multi-source heterogeneous data acquisition and preprocessing module includes: The distributed sensor network unit is deployed at key nodes of the pipeline, including pump station inlets and outlets, valves, bends and straight sections, and includes pressure sensors, temperature sensors, flow meters, acoustic emission sensors and distributed fiber optic sensing systems. The data cleaning unit is configured to use an algorithm based on statistical outlier detection to identify and replace outliers in the original sensor data. The algorithm operates based on a sliding window containing a preset number of neighborhood points and a standard deviation threshold of a preset multiple. The spatiotemporal alignment unit is configured to timestamp all sensor data based on a time synchronization protocol and map the measurement data to a unified spatial coordinate system using a pipeline geometry model. The feature standardization unit is configured to standardize the data processed by the spatiotemporal alignment unit to eliminate differences in physical dimensions and magnitudes.

[0009] Specifically, the physical mechanism constraint embedding module includes: The governing equation construction unit is configured to construct a set of governing equations based on the laws of mass conservation, momentum conservation, and energy conservation, applicable to the transient process of cryogenic multiphase flow of LNG, and taking into account fluid compressibility, latent heat of phase change, and pipe wall friction effects. The equation discretization unit is configured to spatially discretize the set of governing equations in the pipeline computational domain using the finite volume method and temporally discretize them using a time-progression scheme to form discretized residual equations. The loss function construction unit is configured to construct a physical constraint loss term from the sum of squares of the discretized residual equations, which is used to measure the deviation between the neural network prediction results and the physical laws described by the set of governing equations.

[0010] Specifically, the control equation construction unit is also used to explicitly add a volumetric force source term caused by the platform's motion acceleration to the momentum equation of the control equation set, so as to incorporate the driving force effect of the offshore platform's motion on the fluid inside the pipe into the physical constraints; correspondingly, the input data of the deep spatiotemporal fusion prediction model module synchronously includes the temporal data of the platform's motion, and the physical constraint loss term forces the prediction result of the model to satisfy the control equation set including the volumetric force source term.

[0011] Specifically, the deep spatiotemporal fusion prediction model module adopts an encoder-decoder architecture and includes: The spatiotemporal feature encoder is composed of a recurrent neural network for capturing the long-range temporal dependencies of the multidimensional temporal data stream and a one-dimensional convolutional neural network for extracting local spatial features along the pipe axis. The physical information fusion layer, located between the encoder and the decoder, is used during the model training phase to substitute the predicted state corresponding to the deep features extracted by the encoder into the physical constraint loss function, and to add the weighted physical constraint loss term to the model's total loss function to force the model's prediction results to meet physical laws, wherein the weighting coefficient is an adjustable hyperparameter within a preset range. The state decoder and predictor, consisting of a fully connected neural network and a deconvolutional network, is used to decode the feature vectors after fusing physical constraints into flow state tensors covering all computational grid points across the entire pipeline at multiple future time steps.

[0012] Specifically, the real-time prediction and visualization output module includes: The result post-processing unit is configured to perform inverse normalization on the flow state tensor output by the deep spatiotemporal fusion prediction model module, and restore it to engineering data with actual physical units. The graphics rendering engine is configured to generate velocity cloud maps, phase interface position animations, and pressure contour maps in real time within the three-dimensional model of the pipeline based on the engineering data. The intelligent alarm unit is configured to perform real-time analysis of the prediction results based on preset safety rules. When the predicted velocity field, pressure gradient, or gas phase volume fraction meets the alarm conditions, it automatically triggers a graded alarm and generates a structured report containing the abnormal location, severity, and handling suggestions.

[0013] According to a second aspect of the present invention, a method for real-time prediction of flow status in an LNG pipeline transportation system is provided, comprising: S1. Through a distributed sensor network deployed in the LNG pipeline system, raw sensor data is collected synchronously, and the raw sensor data is cleaned, spatiotemporally aligned and standardized preprocessed to form multi-dimensional time-series input data. S2. Construct a deep spatiotemporal fusion prediction model with embedded physical mechanism constraints. The model takes the multidimensional time-series input data as input and introduces a physical constraint loss term obtained by discretization of the LNG multiphase flow control equation during the training process to drive the model to learn a dynamic evolution mode that conforms to physical laws. S3. Use historical working condition data to train the deep spatiotemporal fusion prediction model offline to optimize the model parameters; S4. Input the multi-dimensional time-series input data that has been collected and preprocessed in real time into the trained deep spatiotemporal fusion prediction model for forward inference, and output the prediction results of the velocity field, phase distribution and pressure fluctuation of the entire pipeline in the future prediction time domain. S5. Perform inverse standardization and visualization rendering on the prediction results to generate a dynamic panoramic view of the pipeline flow state and perform real-time analysis and alarm.

[0014] Specifically, the cleaning process employs an algorithm based on statistical outlier detection. A sliding window containing a preset number of neighboring points is constructed with the current data point as the center. If the absolute difference between the value of the current data point and the mean of the data within the window exceeds a standard deviation threshold based on a preset multiple, it is determined to be an outlier and replaced. The alignment process uses a time synchronization protocol to synchronize timestamps and utilizes a pipeline geometry model to map the measurement data to a unified spatial coordinate system. The standardization process performs standardization calculations on data of different physical quantities to eliminate differences in dimensions and magnitudes.

[0015] Specifically, the physical constraint loss term is constructed based on the sum of squares of the residual equations after discretization of the LNG multiphase flow control equations; the total loss function of the deep spatiotemporal fusion prediction model is the sum of the data loss term and the weighted physical constraint loss term, wherein the coefficients used for weighting are adjustable hyperparameters within a preset range. By optimizing the total loss function, the prediction results of the model can simultaneously satisfy the data fitting accuracy and physical law constraints.

[0016] Specifically, the distributed sensor network additionally includes motion attitude sensors for collecting platform motion acceleration data; in the step of constructing a deep spatiotemporal fusion prediction model, the LNG multiphase flow control equation includes a volume force source term dynamically calculated from real-time platform motion acceleration data, and the multidimensional time-series input data synchronously includes time-series data of platform motion, so that the trained model can predict the coupling effect between platform motion and fluid flow in the pipe.

[0017] Beneficial effects: The real-time prediction method and system for the flow state of an LNG pipeline transportation system provided by this invention brings many significant benefits by cleverly integrating data-driven approaches and physical model constraints: First, this invention achieves high-precision, multi-dimensional dynamic reconstruction of the flow state of the entire LNG pipeline, fundamentally overcoming the limitations of traditional discrete-point monitoring methods. By deploying a multi-source heterogeneous sensor network including pressure, temperature, flow, acoustic, and distributed optical fibers, and through rigorous spatiotemporal alignment and preprocessing, the system can acquire synchronous, high-quality data covering key nodes throughout the pipeline. Based on this, the deep spatiotemporal fusion prediction model can output the velocity, phase fraction, and pressure field covering every computational grid point of the pipeline within a future time window, thereby expanding state perception from finite "points" to continuous "fields," greatly improving the comprehensiveness, continuity, and spatial resolution of state monitoring. This enables dispatchers to have a global grasp of flow evolution, laying the foundation for precise control.

[0018] Second, this invention significantly enhances the generalization ability and extrapolation performance of the neural network model by discretizing the control equations describing the nonlinear evolution of LNG multiphase flow into physical constraint loss terms and deeply embedding them into the training process. This method ensures that the model not only pursues historical fitting accuracy to the training data during optimization but also forces its predictions to satisfy the laws of conservation of mass, momentum, and energy. This soft constraint of physical laws is equivalent to injecting prior knowledge into the data-driven model, guiding it to learn dynamic evolution patterns that conform to physical essence, rather than superficial data correlations. Therefore, even under unseen operating conditions not fully covered by the training data (such as more drastic flow regulation, rapid start-up and shutdown, or sudden external disturbances), the model can make more reasonable predictions based on physical consistency. This effectively solves the key problems of weak generalization and potential failure in unknown scenarios of purely data-driven methods, improving the system's robustness in dealing with complex and variable operating conditions.

[0019] Third, this invention possesses millisecond-level data processing and model inference capabilities, enabling real-time response and advanced prediction of rapidly changing operating conditions, providing a valuable decision-making window for safe dispatching. Through an optimized data pipeline, lightweight model architecture, and dedicated computing hardware, the system achieves end-to-end low latency from data acquisition to prediction result output. This allows the system to predict risks such as sudden pressure increases and slug flow formation that may occur in the pipeline in the near real-time within a short period (e.g., the next tens of minutes), thus transforming traditional "post-event alarms" into "pre-event warnings." Dispatchers can take intervention measures in advance based on the predicted information, such as adjusting pump frequency or valve opening, effectively avoiding or mitigating accidents and significantly improving the safety and operational stability of the LNG pipeline transportation system in the face of sudden disturbances.

[0020] Fourth, this invention reduces reliance on manual analysis and subjective errors through end-to-end automated data processing, model prediction, visualization, and intelligent alarm processes, thus promoting the intelligentization and automation of pipeline operation and maintenance. The system automatically completes the entire transformation from raw signals to decision support information and presents the results in intuitive graphics, animations, and structured reports, significantly reducing the cognitive load on dispatchers. This not only improves operation and maintenance efficiency but also provides core, highly reliable state prediction data support for subsequent advanced applications such as intelligent operation and maintenance, energy efficiency optimization, and leak early warning, helping the entire transportation system move towards a higher level of autonomous and intelligent operation and maintenance. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall technical architecture of a real-time prediction method and system for the flow state of an LNG pipeline transportation system proposed in a specific embodiment of the present invention. Figure 2 This is a schematic diagram of the core principle framework of the deep spatiotemporal fusion prediction model module proposed in a specific embodiment of the present invention; Figure 3 This is a logical flow diagram of the multi-source heterogeneous data acquisition and preprocessing module proposed in a specific embodiment of the present invention; Figure 4 This is a schematic diagram of the principle framework of the physical mechanism constraint embedding module proposed in a specific embodiment of the present invention; Figure 5 This is a schematic diagram of the logic flow and interaction of the real-time prediction and visualization output module proposed in a specific embodiment of the present invention; Figure 6 This is a flowchart illustrating the real-time prediction method for the flow state of an LNG pipeline transportation system proposed in a specific embodiment of the present invention. Detailed Implementation

[0022] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Based on the embodiments in this application, other similar embodiments obtained by those skilled in the art without creative effort should all fall within the scope of protection of this application. Furthermore, directional terms mentioned in the following embodiments, such as "up," "down," "left," and "right," are only for reference to the directions in the accompanying drawings; therefore, the directional terms used are for illustrative purposes and not for limiting the invention.

[0023] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments.

[0024] Please see Figures 1-5 This invention provides a real-time flow state prediction system for an LNG pipeline transportation system, comprising: The multi-source heterogeneous data acquisition and preprocessing module is used to acquire raw sensor data from a distributed sensor network and perform cleaning, spatiotemporal alignment and standardization to generate a multi-dimensional time-series data stream. The physical mechanism constraint embedding module is used to construct a set of control equations describing the multiphase flow process in LNG pipelines and discretize them into physical constraint loss functions that can be embedded in neural network training. The deep spatiotemporal fusion prediction model module is used to receive the multidimensional time series data stream, and perform training and inference based on the physical constraint loss function to generate prediction results of the velocity field, phase distribution and pressure fluctuation of the entire pipeline within the future time window; The real-time prediction and visualization output module is used to perform inverse normalization and visualization rendering on the prediction results and generate alarm information.

[0025] It should be further explained that the system is an end-to-end integrated software and hardware solution. At the hardware level, it relies on a distributed sensor network deployed on the LNG pipeline physical system, as well as servers and industrial control computers for data computation, storage, and visualization. At the software level, the four modules work collaboratively to form a complete closed loop, from raw data perception to intelligent prediction through physical law fusion, and finally to decision support information output. The multi-source heterogeneous data acquisition and preprocessing module is responsible for communicating with field sensors, acquiring raw data streams, and performing normalization processing to provide high-quality input for subsequent models. The physical mechanism constraint embedding module operates independently of the real-time data stream. It constructs a mathematical model framework based on the fundamental principles of fluid mechanics and transforms this framework into a computable loss function to constrain and guide the data-driven model's learning process. The deep spatiotemporal fusion prediction model module is the intelligent core of the entire system. It receives preprocessed real-time data and uses an offline-trained neural network model embedded with physical constraints to perform millisecond-level forward computation, outputting a comprehensive prediction of future flow states. The real-time prediction and visualization output module transforms the numerical tensors output by the model into graphical, animated, and structured alarm information that dispatchers can intuitively understand, thereby providing advanced decision support for on-site operations.

[0026] Understandably, this system achieves high-precision, multi-dimensional dynamic reconstruction of the flow state across the entire LNG pipeline by integrating distributed sensor data with deep physical mechanisms, overcoming the limitations of traditional point-based monitoring. By embedding the control equations describing the nonlinear evolution of multiphase flow as soft constraints into neural network training, the system significantly enhances the model's generalization ability and extrapolation performance under unseen operating conditions (such as rapid start-up and shutdown, and external disturbances), solving the problem of weak generalization in purely data-driven methods. The system possesses millisecond-level end-to-end processing and inference capabilities, enabling real-time responses to transient operating conditions and providing advanced warnings for safe scheduling, effectively improving the safety and operational stability of the LNG pipeline transportation system.

[0027] Specifically, the multi-source heterogeneous data acquisition and preprocessing module includes: The distributed sensor network unit is deployed at key nodes of the pipeline, including pump station inlets and outlets, valves, bends and straight sections, and includes pressure sensors, temperature sensors, flow meters, acoustic emission sensors and distributed fiber optic sensing systems. The data cleaning unit is configured to use an algorithm based on statistical outlier detection to identify and replace outliers in the original sensor data. The algorithm operates based on a sliding window containing a preset number of neighborhood points and a standard deviation threshold of a preset multiple. The spatiotemporal alignment unit is configured to timestamp all sensor data based on a time synchronization protocol and map the measurement data to a unified spatial coordinate system using a pipeline geometry model. The feature standardization unit is configured to standardize the data processed by the spatiotemporal alignment unit to eliminate differences in physical dimensions and magnitudes.

[0028] It should be further explained that the distributed sensor network units constitute the "sensory nerves" of the system. Their deployment on the pipeline follows these principles: (1) Pressure and temperature sensors are installed in pairs and arranged at preset intervals (e.g., 200 meters) in straight pipe sections to monitor the basic thermal parameters of the fluid. High-frequency dynamic pressure sensors are added near key equipment, such as 50 meters before the inlet and 50 meters after the outlet of the pump station, 1 meter before and after each main valve, and on the outside of all bends greater than 30 degrees, to capture pressure pulsations caused by drastic changes in fluid momentum.

[0029] (2) Ultrasonic flow meters are preferred and are installed at the outlet of each pumping station and the inlet of the distribution station to provide volumetric flow data of pipeline transportation.

[0030] (3) The acoustic emission sensor array is deployed around potential weak links such as pipe welds and flange connections to collect broadband acoustic signals generated by fluid cavitation, phase change or micro-leakage. These signals are important basis for judging local flow state and structural health.

[0031] (4) A distributed optical fiber sensing system is laid in parallel along the entire pipeline. Based on the principles of Raman scattering and Brillouin scattering, it can continuously measure the temperature field and strain field distribution along the pipeline axis with high spatial resolution (e.g., 1 meter), thereby realizing continuous spatial perception of the physical state of the pipeline.

[0032] Furthermore, the data cleaning unit employs an algorithm based on statistical outlier detection. The preferred implementation steps of this algorithm are as follows: For each sensor data stream, a sliding window is formed by taking a certain number of points forward and backward from the current data point. Preferably, the number of neighboring points (k) within the window is set to 50. The mean (μ) and standard deviation (σ) of all data within the window are calculated. If the absolute difference between the current data point's value (x) and the mean μ exceeds a preset multiple of the standard deviation, it is determined to be an outlier. Preferably, this multiple (threshold) is set to 2.0. The reason for choosing k=50 is that it provides a sufficiently large statistical sample to stably calculate the mean and standard deviation, while avoiding a window that is too long, leading to a delayed response to rapid changes. Choosing threshold=2.0 is based on the assumption of a normal distribution; approximately 95% of the data points fall within the range of the mean ± 2 times the standard deviation, and points outside this range have a high probability of being outliers. The outliers are replaced by a replacement value generated by an algorithm based on linear interpolation and historical trend extrapolation to maintain the continuity of the data sequence.

[0033] Furthermore, the spatiotemporal alignment unit first assigns a uniform, millisecond-level timestamp to each data packet based on a high-precision network time protocol. Then, using the 3D geometric model of the pipeline stored in the system (including pipeline direction, elevation, diameter, material, and precise coordinates of all components), the measurements from each discrete sensor are mapped to the corresponding 3D spatial coordinates of the model. For continuous spatial data from distributed optical fibers, it is directly registered with the model according to its inherent resolution. This ultimately generates a spatiotemporally strictly synchronized multidimensional data field.

[0034] Furthermore, the feature standardization unit performs alignment on the physical quantity data (such as pressure P, temperature T, flow rate Q, acoustic emission signal intensity A, and fiber temperature). Fiber strain Z-score standardization is performed on each of the following: (e.g., ...). For a given sequence of physical quantities... Its standardized value The calculation is as follows: ,in This is the average value of the physical quantity over a long period of historical data. Its standard deviation. After this processing, all data are transformed into a distribution with a mean of 0 and a standard deviation of 1, eliminating the influence of units and magnitude on model training.

[0035] Understandably, this module, through the construction of a comprehensive sensor network and a rigorous data preprocessing process, provides high-quality, spatiotemporally aligned, and dimensionless regularized input data for subsequent models, which is the foundation for achieving high-precision predictions. The use of statistical outlier detection and interpolation algorithms effectively filters out noise caused by electromagnetic interference, sensor malfunctions, and other factors, improving data quality. Millisecond-level time synchronization and three-dimensional spatial mapping ensure the consistency and comparability of data from different locations and types of sensors in the spatiotemporal dimensions, providing data support for the subsequent construction of the flow field across the entire pipe section. Z-score normalization accelerates the convergence process of the neural network model and avoids training instability caused by excessive differences in the scale of input features.

[0036] Specifically, the physical mechanism constraint embedding module includes: The governing equation construction unit is configured to construct a set of governing equations based on the laws of mass conservation, momentum conservation, and energy conservation, applicable to the transient process of cryogenic multiphase flow of LNG, and taking into account fluid compressibility, latent heat of phase change, and pipe wall friction effects. The equation discretization unit is configured to spatially discretize the set of governing equations in the pipeline computational domain using the finite volume method and temporally discretize them using a time-progression scheme to form discretized residual equations. The loss function construction unit is configured to construct a physical constraint loss term from the sum of squares of the discretized residual equations, which is used to measure the deviation between the neural network prediction results and the physical laws described by the set of governing equations.

[0037] It should be further explained that the governing equation construction unit is based on the fundamental laws of fluid mechanics, constructing a set of governing equations applicable to the cryogenic, compressible, gas-liquid two-phase transient process of LNG. This set of equations is a coupled system, with fluid velocity as its core variable. ,pressure ,temperature Gas phase volume fraction and density The equations explicitly consider the key physical effects of LNG transported via pipeline: (1) The compressibility of fluids, i.e., the relationship between density and pressure and temperature. .

[0038] (2) The latent heat absorption and release that accompany the gas-liquid phase change process is reflected in the energy equation through the source term.

[0039] (3) The resistance effect of pipe wall friction on flow is usually introduced into the momentum equation through empirical relations such as the Darcy-Weisbach formula.

[0040] The discretization of the equations employs the mature finite volume method from computational fluid dynamics. First, the entire pipeline computational domain (which may be a one-dimensional, two-dimensional, or simplified three-dimensional model) is divided into numerous non-overlapping control volumes (grids). On each control volume, the control equations in partial differential form are integrated, and the Gaussian divergence theorem is applied to transform the volume integral into a surface integral on the control volume surface, thus obtaining discrete equations describing the conservation of physical quantities on each control volume. For the time dimension, an implicit time-progression scheme (such as the backward Euler method) with good numerical stability is used for discretization. After discretization, the original system of partial differential equations is transformed into a large system of nonlinear algebraic equations concerning all state variables at all grid points.

[0041] The core task of the loss function construction unit is to construct the physical constraint loss term. The principle is: the neural network model predicts the overall state of the pipeline at a certain moment (i.e., the state of each grid point). , , , Substitute these values ​​into the aforementioned discretized system of algebraic equations. Since the predicted state is not an exact solution to the equations, the left and right sides of the equations will not be completely equal after substitution, resulting in a residual. The physical constraint loss term is defined as the sum of the squared residuals across all computational units. Its mathematical expression is: .in, The residual operator represents the discretized set of governing equations; The model predicts the first The set of state variables for each computing unit; These are the parameters (weights and biases) of the neural network model. The total number of computational units divided into the pipeline computational domain; This represents the L2 norm of the vector. The smaller the value of this loss term, the more the predictions of the neural network conform to the physical laws described by the conservation of mass, momentum, and energy.

[0042] Understandably, this module introduces first-principles physics into the data-driven model in the form of a computable loss function. By using the residuals of the governing equations as optimization terms, the model is guided during training to not only fit the data but also adhere to fundamental physical conservation laws. This is equivalent to providing the model with a "physical mentor," greatly enhancing the physical rationality and consistency of the model's predictions. Especially for complex processes like LNG multiphase flow, which are highly nonlinear and involve phase transitions, pure data models are prone to producing absurd predictions under conditions not covered by the training data. The embedding of physical constraints effectively constrains the model's output space, significantly improving the model's generalization ability and reliability under unseen conditions (extrapolation). This is one of the core innovations of this invention.

[0043] Specifically, the control equation construction unit is also used to explicitly add a volumetric force source term caused by the platform's motion acceleration to the momentum equation of the control equation set, so as to incorporate the driving force effect of the offshore platform's motion on the fluid inside the pipe into the physical constraints; correspondingly, the input data of the deep spatiotemporal fusion prediction model module synchronously includes the temporal data of the platform's motion, and the physical constraint loss term forces the prediction result of the model to satisfy the control equation set including the volumetric force source term.

[0044] It should be further noted that this embodiment specifically considers the application scenario of offshore floating liquefied natural gas (LNG) units (such as FLNG). In this scenario, the entire pipeline system moves (rolls, pitches, heaves) along with the offshore platform. The platform's acceleration generates additional volume forces on the fluid within the pipeline (similar to inertial forces in a non-inertial frame of reference), significantly affecting the fluid's momentum distribution and potentially inducing specific flow instabilities.

[0045] Therefore, the governing equation building unit explicitly adds a volume force source term caused by the platform's motion acceleration to the right side of the standard governing equation set (especially the momentum conservation equation). This source item is based on real-time collected platform motion acceleration data (such as triaxial acceleration). Dynamic calculation. For example, considering gravity, the volumetric force source term can be expressed as... ,in For fluid density, For gravitational acceleration. This additional term directly incorporates the driving force effect of the platform motion on the fluid inside the pipe into the mathematical description of the controlling physical process.

[0046] Accordingly, at the input end of the deep spatiotemporal fusion prediction model module, in addition to the sensing data of the pipeline fluid itself (pressure, temperature, etc.), the time-series data of the platform motion (such as roll angle, pitch angle, and heave acceleration) also need to be input synchronously. This enables the model to sense external motion stimuli.

[0047] More importantly, during the model training phase, the physical constraint loss term is calculated. At that time, the governing equations used included the aforementioned volume force source terms. The system of equations. Therefore, the physical constraint loss term. The model's predictions must satisfy the fluid motion laws in a wobbling coordinate system. This allows the model to learn and predict the coupled dynamics between the platform's motion and the complex fluid flow within the pipe.

[0048] Understandably, this extension enables the system and method of this invention to be directly applied to LNG pipeline systems in marine environments characterized by swaying motion. By using platform acceleration as the source term in the physical equations and the input to the model, the system can accurately model and predict unique marine conditions such as "slug flow forming at the low point of the pipeline during a specific roll cycle" or "the impact of sudden platform heave on pump outlet pressure." This solves the problem that traditional land-based pipeline models or models that do not consider motion coupling cannot be directly applied to marine environments, greatly expanding the application scope of this invention and improving its prediction accuracy and reliability in complex marine environments.

[0049] Specifically, the deep spatiotemporal fusion prediction model module adopts an encoder-decoder architecture and includes: The spatiotemporal feature encoder is composed of a recurrent neural network for capturing the long-range temporal dependencies of the multidimensional temporal data stream and a one-dimensional convolutional neural network for extracting local spatial features along the pipe axis. The physical information fusion layer, located between the encoder and the decoder, is used during the model training phase to substitute the predicted state corresponding to the deep features extracted by the encoder into the physical constraint loss function, and to add the weighted physical constraint loss term to the model's total loss function to force the model's prediction results to meet physical laws, wherein the weighting coefficient is an adjustable hyperparameter within a preset range. The state decoder and predictor, consisting of a fully connected neural network and a deconvolutional network, is used to decode the feature vectors after fusing physical constraints into flow state tensors covering all computational grid points across the entire pipeline at multiple future time steps.

[0050] It should be further explained that the deep spatiotemporal fusion prediction model module is the intelligent core of the system, and its structure and workflow are as follows: (1) Spatiotemporal Feature Encoder: It is responsible for extracting deep spatiotemporal features from preprocessed multidimensional time-series data. It is usually composed of two cascaded network layers. The first layer is a gated recurrent unit network, which processes the input time-series data in chronological order. Through update gate and reset gate mechanisms, the GRU can effectively capture long-range temporal dependencies in the data that span long time steps (e.g., hundreds of time steps), such as identifying the propagation period of pressure waves, the slow trend of temperature changes, or the delayed effect of flow changes. The second layer is a one-dimensional convolutional neural network, whose convolution kernels perform sliding convolution operations in the spatial dimension of the pipe axis. The one-dimensional CNN can extract spatial local correlation features between sensor data at adjacent locations in the pipe, such as identifying upstream and downstream pressure change patterns caused by valve throttling, flow pattern changes caused by topographic relief, or spatial anomalies in the temperature field caused by local heating.

[0051] (2) Physical Information Fusion Layer: This is the key layer for realizing the embedding of physical constraints. During the offline training phase of the model, this layer performs the following operations: First, the deep features extracted by the encoder are passed through a forward propagation process to obtain the model's predicted state of the flow field at the current or future time (i.e., the velocity, pressure, phase fraction, etc. of each grid point). Then, this predicted state is... and current parameters of the model Substitute the physical constraint loss function constructed by the physical mechanism constraint embedding module In this process, the degree to which the current prediction violates the laws of physics is calculated. Then, in the overall objective function of model optimization... In addition to the data loss term that measures the difference between the predicted value and the true label, (e.g., mean squared error), forcibly add a weighted physical constraint loss term. That is, the total loss function is: .in, It is an adjustable hyperparameter used to balance the weight between the accuracy of data fitting and the degree to which physical laws are satisfied. Through the backpropagation algorithm, the model parameters... The optimization process not only aims to reduce prediction error The update is directed towards reducing the residuals of the physical equations. The direction is updated. Preferably, after grid search, The value is set between 0.1 and 1.0, with a preferred empirical value of 0.5. (Select) The reason is that it can fit the training data without excessively interfering with the model's ability to fit the training data. Under the premise of (dominance), apply sufficiently strong physical laws to guide ( (Effective), achieving a good balance between prediction accuracy and physical consistency in practice.

[0052] (3) State Decoder and Predictor: It is responsible for mapping the feature vectors, which incorporate physical constraint information, back to the specific flow state prediction space. First, a fully connected neural network maps the feature vectors to a high-dimensional latent space representation. Then, a deconvolutional network performs the opposite operation to the one-dimensional CNN in the encoder, gradually restoring (decoding) the compressed feature representation to the target spatial-temporal resolution through multiple upsampling and deconvolution (or transposed convolution) operations. Finally, the decoder outputs a three-dimensional tensor, whose three dimensions correspond to: the number of grid points along the pipe axis (e.g., N), the time step of future prediction (e.g., T time steps), and the number of channels in the flow state (e.g., C=3, corresponding to axial velocity, gas volume fraction, and static pressure, respectively). In this way, the model completes the high-dimensional mapping from historical multi-source time series data to the future full pipe section, multiple time points, and multiple physical quantity fields.

[0053] Understandably, this module effectively addresses the problem of spatiotemporal sequence prediction through an encoder-decoder architecture. The cascaded design of GRU and one-dimensional CNN excels at capturing temporal dependencies and spatial features, respectively, making it highly suitable for one-dimensional topologies like pipelines with strong spatiotemporal correlations. The design of the physical information fusion layer is the essence of this invention. By using the residuals of the physical equations as a loss term, it "teaches" prior physical knowledge to the data-driven neural network in a soft-constraint manner. This ensures that the dynamic evolution patterns learned by the model not only conform to the statistical regularities of historical data but also strictly follow the fundamental laws of fluid mechanics, thus achieving "physics-guided machine learning." This fusion greatly enhances the model's predictive ability and extrapolation robustness under extreme or transient conditions not covered by the training data. The decoder ultimately outputs complete field information, realizing the leap from "point measurement" to "field prediction" and providing the possibility for comprehensive state awareness.

[0054] Specifically, the real-time prediction and visualization output module includes: The result post-processing unit is configured to perform inverse normalization on the flow state tensor output by the deep spatiotemporal fusion prediction model module, and restore it to engineering data with actual physical units. The graphics rendering engine is configured to generate velocity cloud maps, phase interface position animations, and pressure contour maps in real time within the three-dimensional model of the pipeline based on the engineering data. The intelligent alarm unit is configured to perform real-time analysis of the prediction results based on preset safety rules. When the predicted velocity field, pressure gradient, or gas phase volume fraction meets the alarm conditions, it automatically triggers a graded alarm and generates a structured report containing the abnormal location, severity, and handling suggestions.

[0055] It should be further explained that this module is responsible for transforming the model's numerical prediction results into information that schedulers can intuitively understand and use for operational decisions. The specific design includes: (1) The post-processing unit receives the flow state tensor output by the model. Since the model input data has been Z-score standardized, its output is also the standardized value. The post-processing unit calls the historical average values ​​of each physical quantity stored in the feature standardization unit. and standard deviation Perform the inverse operation on each data point in the tensor: This allows the predicted values ​​to be converted into engineering data with actual physical units, such as velocity in meters per second (m / s), pressure in megapascals (MPa), and gas volume fraction in percentage (%).

[0056] (2) The graphics rendering engine drives the visualization interface based on the post-processed engineering data. The engine has a built-in 3D geometric model of the pipeline that is consistent with the preprocessing module. Its workflow is as follows: a) For the velocity field, the engine maps the flow velocity magnitude of each grid point (or segment) in the pipeline to a specific color spectrum (e.g., red represents high speed and blue represents low speed), and renders a velocity cloud map in real time on the surface or longitudinal section of the pipeline 3D model, dynamically displaying the distribution and evolution of the flow velocity along the pipeline.

[0057] b) For phase distribution, the engine calculates the gas phase volume fraction at each grid point. Value, setting a transparency or color threshold (e.g., The area is rendered as a translucent light color representing gas. The area is rendered as an opaque dark color representing the liquid, thereby generating a dynamic animation of the movement of the gas-liquid two-phase interface in the pipe and the formation and development of slug flow.

[0058] c) For pressure fluctuations, the engine plots pressure contour lines on the pipeline model or generates pressure distribution curves along the pipeline axis, visually displaying the location of high-pressure and low-pressure areas and the strength of pressure gradients. All these graphical elements are displayed panoramically on a large screen in the dispatch center at a high refresh rate (e.g., 10 frames per second or higher), providing operators with a global and intuitive understanding of the flow status.

[0059] (3) The intelligent alarm unit performs real-time, automatic security scanning and analysis of the restored engineering data. It has pre-set static and dynamic threshold rules for a series of key parameters, such as: a) Pressure alarm rules: The instantaneous pressure value at any point in the pipeline shall not exceed 1.1 times the design pressure (absolute threshold); the rate of change of pressure within 1 second shall not exceed 2 MPa / s (dynamic gradient threshold).

[0060] b) Flow pattern risk rule: If the gas phase volume fraction exceeds 0.8 in a local continuous 1-meter pipe section and lasts for more than 10 seconds, it is judged as a gas lock risk warning.

[0061] c) Flow Safety Rules: A velocity difference exceeding 5 m / s between adjacent grid points is considered a severe shear flow warning. This unit continuously scans the prediction data for each grid point and each time step within the future prediction time window (e.g., the next 30 minutes). Once any parameter combination triggers the above rules, a tiered alarm is immediately triggered (e.g., yellow warning, orange alarm, red emergency alarm). Simultaneously, the unit automatically generates a structured report containing: the pipeline mileage location where the anomaly is predicted to occur (e.g., KP85+200), the estimated time of occurrence (e.g., 15 minutes later), the anomaly type (e.g., pressure surge), and the severity level. Based on a built-in knowledge base or rule base, preliminary handling suggestions are provided (e.g., "A pressure surge is predicted at KP85+200 in 15 minutes; it is recommended to reduce the output power of upstream pump station 2 in advance"). Alarms and reports are pushed to relevant personnel via the dispatch system's audio-visual devices, screen pop-ups, SMS, or work order system.

[0062] Understandably, this module transforms complex numerical prediction results into intuitive visual information and clear action instructions, which is key to human-computer interaction. Denormalization processing gives the prediction results direct engineering significance. The dynamic, visualized panoramic view allows dispatchers to grasp the evolution trend of the flow status across the entire pipeline segment at a glance, surpassing the limitations of traditional monitoring systems that only display discrete point data. The intelligent alarm unit realizes the transformation from "post-event alarm" to "pre-event warning." Through predictive analysis of future states, it issues alerts and provides handling suggestions before potential risks actually occur, providing dispatchers with a valuable window for proactive decision-making, greatly improving the safety and stability of the pipeline system in responding to transient conditions and emergencies.

[0063] Please see Figure 6 The present invention provides another embodiment, which provides a method for real-time prediction of the flow state of an LNG pipeline transportation system. The method for real-time prediction of the flow state of an LNG pipeline transportation system includes: S1. Through a distributed sensor network deployed in the LNG pipeline system, raw sensor data is collected synchronously, and the raw sensor data is cleaned, spatiotemporally aligned and standardized preprocessed to form multidimensional time-series input data.

[0064] It should be further explained that this invention provides a systematic, end-to-end process, covering the entire chain from data preparation, model building and training, to online prediction and application. First, comprehensive state information of the pipeline system is acquired synchronously through a high-density, multi-type sensor network, and then rigorously preprocessed to eliminate noise, align with spatiotemporal references, and unify data scales, providing high-quality input for subsequent analysis.

[0065] S2. Construct a deep spatiotemporal fusion prediction model with embedded physical mechanism constraints. The model takes the multidimensional time-series input data as input and introduces a physical constraint loss term obtained by discretization of the LNG multiphase flow control equation during the training process to drive the model to learn a dynamic evolution mode that conforms to physical laws.

[0066] It should be further explained that the core of this invention is to construct a deep spatiotemporal fusion prediction model. Its unique feature lies in adding the residuals formed after discretizing the governing equations describing the physical laws of LNG multiphase flow as part of the loss function into the training objective of the neural network model. This forces the model to learn statistical laws from the data while also requiring its prediction results to satisfy the laws of conservation of mass, momentum, and energy as much as possible, thus endowing the model with strong physical consistency and extrapolation capabilities.

[0067] S3. Use historical operating data to train the deep spatiotemporal fusion prediction model offline to optimize the model parameters.

[0068] It should be further explained that this invention utilizes accumulated historical normal operating condition data (which may come from SCADA systems, high-precision simulations, or experiments) to fully train the model offline, optimizing its internal parameters until a preset prediction accuracy (e.g., mean absolute error below 2.5%) is achieved on an independent validation set. The trained model can then be deployed online.

[0069] S4. Input the multi-dimensional time-series input data, which has been collected and preprocessed in real time, into the trained deep spatiotemporal fusion prediction model for forward inference, and output the prediction results of the velocity field, phase distribution and pressure fluctuation of the entire pipeline in the future prediction time domain.

[0070] It should be further explained that during the online application phase, the system continuously inputs the pre-processed data into the model in real time using a sliding window approach (for example, using data from the past 30 minutes to predict the next 10 minutes). The model performs efficient forward inference calculations (the time for a single inference can be less than 50 milliseconds) and outputs a complete prediction of the future flow state of the entire pipe segment.

[0071] S5. Perform inverse standardization and visualization rendering on the prediction results to generate a dynamic panoramic view of the pipeline flow state and perform real-time analysis and alarm.

[0072] It should be further explained that the present invention restores these numerical prediction results to engineering values ​​through inverse standardization, drives the visualization system to generate a dynamic panoramic view, and at the same time, the intelligent alarm module performs real-time security scanning on the prediction results, and issues an early warning once a potential risk is detected.

[0073] Understandably, this method, through a systematic approach, deeply integrates data-driven methods with physical models, achieving real-time, high-precision, and comprehensive prediction of complex flow states in LNG pipelines. Compared to methods relying on simplified models or purely data-driven approaches, this method's predictions are more consistent with physical reality and exhibit stronger generalization capabilities. Compared to high-fidelity computational fluid dynamics real-time simulation, this method, while maintaining physical consistency, achieves millisecond-level prediction speeds through neural network forward inference, meeting the real-time requirements of industrial sites for early warning of rapid transient processes (such as slug flow and water hammer). The entire process is highly automated, reducing reliance on human experience and providing core technical support for intelligent operation and maintenance, safe scheduling, and energy efficiency optimization of LNG pipelines.

[0074] Specifically, the cleaning process employs an algorithm based on statistical outlier detection. A sliding window containing a preset number of neighboring points is constructed with the current data point as the center. If the absolute difference between the value of the current data point and the mean of the data within the window exceeds a standard deviation threshold based on a preset multiple, it is determined to be an outlier and replaced. The alignment process uses a time synchronization protocol to synchronize timestamps and utilizes a pipeline geometry model to map the measurement data to a unified spatial coordinate system. The standardization process performs standardization calculations on data of different physical quantities to eliminate differences in dimensions and magnitudes.

[0075] It should be further explained that the preprocessing step is crucial to ensuring the quality of the input data for the prediction model, and the specific design is as follows: (1) Cleaning process: In the specific implementation of the algorithm based on statistical outlier detection, for the data stream of each sensor At any time With the current data point Take the center and move forward. A historical point, looking backward A point (or the most recent one from history) (a number of points), forming a collection Sliding window of neighborhood points ,in Preferably, a symmetrical window is used. ,Right now (Including the center point), or take the 50 most recent historical points ( Calculate the mean of all data within this window. and standard deviation Set a threshold multiple. (Preferred value: 2.0). If satisfied... Then determine These are outliers. Outliers that have been removed are replaced using linear interpolation or extrapolation based on the historical data trends of the sensor to maintain the continuity of the data sequence.

[0076] (2) Alignment process: Time synchronization is achieved based on a high-precision network time protocol, ensuring that the timestamp deviation of all sensor data is within milliseconds. Spatial alignment utilizes the three-dimensional geometric model of the pipeline stored in the system's digital twin model. This model accurately records the three-dimensional coordinates of the pipeline centerline, pipe diameter, elevation, and the positions of all components (valves, pumps, elbows). The installation location of each discrete sensor (e.g., mileage marker) is mapped to the corresponding spatial coordinates of this three-dimensional model. The measured value represents the state of that coordinate point. For continuous measurements of distributed optical fibers, each sampling point is automatically assigned a series of spatial coordinates based on the fiber optic laying path and sampling interval.

[0077] (3) Standardization process: for each physical quantity (such as pressure) ,temperature Given a real-time data sequence (etc.), perform the following calculations: .in, and These are the global mean and global standard deviation of the physical quantity calculated from historical data over a long period (e.g., the past year). For newly commissioned pipelines without historical data, the design operating conditions or statistical values ​​from short-term operating data can be used for initialization, and these values ​​can be continuously updated as the pipeline operates.

[0078] Understandably, this preprocessing step is crucial. Effective cleaning removes noise and outliers introduced by sensor malfunctions, electromagnetic interference, etc., ensuring the reliability of the input data. Strict spatiotemporal alignment unifies data from different sources, times, and locations under the same reference frame, which is a prerequisite for subsequent full-segment "field" reconstruction and spatiotemporal correlation analysis. Standardization normalizes all features to a similar numerical range, avoiding problems such as unstable gradient updates or certain features being ignored by the model during neural network training due to large differences in the dimensions and magnitudes of physical quantities such as pressure (MPa level), temperature (K level), and flow rate. This significantly accelerates model convergence and improves the final performance.

[0079] Specifically, the physical constraint loss term is constructed based on the sum of squares of the residual equations after discretization of the LNG multiphase flow control equations; the total loss function of the deep spatiotemporal fusion prediction model is the sum of the data loss term and the weighted physical constraint loss term, wherein the coefficients used for weighting are adjustable hyperparameters within a preset range. By optimizing the total loss function, the prediction results of the model can simultaneously satisfy the data fitting accuracy and physical law constraints.

[0080] It should be further explained that the embedding of physical constraints is achieved by adding a dedicated loss term to the objective function of the model training. This is the core of this method, and the specific design is as follows: (1) Physical constraint loss term The specific structure has been detailed above; its essence is to measure the flow field state predicted by the neural network. (from model parameters) (Determines) the extent to which the discretized set of fluid dynamics governing equations is satisfied. Its calculation requires first determining the predicted state. Substituting parameters such as velocity, pressure, and phase fraction into the discretized control equations, the calculations for each computational unit are obtained. residual Then find all Sum of squares of the L2 norm of the residuals of each unit: The smaller this value, the more the prediction conforms to the laws of physics.

[0081] (2) Overall objective function for model training It is a weighted sum of data-driven loss and physical constraint loss: . It is a data loss term, usually using the error between the predicted value and the true label in the training data, such as mean squared error or mean absolute error. It forces the model to learn the mapping relationship from the input data to the output flow field. It is a physical constraint loss term. It is an adjustable hyperparameter between 0.1 and 1.0, used to balance the relative importance of these two losses. When When the model degenerates into a purely data-driven model, when When the value is too large, it may over-constrain the model, making it difficult to fit the training data. This can be addressed through cross-validation or experience. This is a good starting point. The optimization process (such as using the Adam optimizer) continuously adjusts the model parameters using gradient descent. To minimize the total loss This means that the optimization process simultaneously pursues two goals: the prediction result should be close to the true value in the training data ( Small), while the prediction results should satisfy the physical equations as much as possible ( Small).

[0082] Understandably, in this way, physical knowledge is no longer something independent of the data model, but is seamlessly and differentiably integrated into the training framework of the machine learning model. The model is no longer merely a "black box" function approximator, but a learner "guided" by physical laws. The beneficial effects of this training strategy are profound: First, it significantly improves the model's generalization ability. Even under new operating conditions not covered by the training data (such as more drastic flow changes or different inlet conditions), due to the constraints of physical conservation laws, the model's predictions will not deviate too far from physical reality, thus possessing extrapolation capabilities. Second, it reduces the dependence on massive, high-precision labeled data (true values ​​of the flow field across the entire pipe section) to some extent, because physical laws themselves provide a powerful regularization. Finally, it makes the model's predictions inherently physically consistent; for example, mass is conserved, and the pressure wave propagation speed is reasonable, which increases the interpretability of the predictions and the engineers' confidence in them.

[0083] Specifically, the distributed sensor network additionally includes motion attitude sensors for collecting platform motion acceleration data; in the step of constructing a deep spatiotemporal fusion prediction model, the LNG multiphase flow control equation includes a volume force source term dynamically calculated from real-time platform motion acceleration data, and the multidimensional time-series input data synchronously includes time-series data of platform motion, so that the trained model can predict the coupling effect between platform motion and fluid flow in the pipe.

[0084] It should be further noted that this is an important extended implementation of the method of the present invention in the application scenario of offshore floating LNG units, and the specific design includes: (1) At the data acquisition end, in addition to sensors (pressure, temperature, flow rate, acoustic emission, distributed optical fiber) that monitor the fluid in the pipeline itself, high-precision motion attitude sensors are installed on the offshore platform (such as the FLNG hull or platform) to collect the platform's six degrees of freedom motion data in real time, especially the angular velocity and linear acceleration of roll, pitch, and heave. These motion data serve as additional input channels and are collected, cleaned, aligned, and standardized synchronously with the fluid sensing data.

[0085] (2) At the level of physical modeling, when constructing the governing equations, add a volume force source term related to the platform acceleration to the standard momentum conservation equation. For a pipe fixed on a platform, in a non-inertial coordinate system, the volume forces acting on the fluid element include gravity and inertial forces. Therefore, the volume force term in the momentum equation can be written as: .in, It is the fluid density. It is gravitational acceleration. It is the platform's translational speed. It is the platform angular velocity. It is the position vector of the fluid element relative to the platform reference point. It is the velocity of the fluid relative to the platform. This is the time derivative. In practical applications, it can be appropriately simplified according to the accuracy and requirements of motion measurement, for example, mainly considering the effects of the platform's linear acceleration and angular acceleration. This time-varying source term directly incorporates the driving effect of the platform's motion on the fluid inside the pipe into the mathematical description of the controlling physical process.

[0086] (3) During model training and application, the multidimensional time-series input data includes not only pipeline pressure and temperature, but also time-series data of platform motion (such as triaxial acceleration and angular velocity). During the training phase, the physical constraint loss is calculated. At that time, the governing equations used included the aforementioned volume force source terms. The complete system of equations. Therefore, the model is in optimization. In the process, it not only learns the internal correlations of fluid data, but also the coupling relationship between fluid data and platform motion data, and its predictions are forced to meet the fluid motion laws in a swaying non-inertial frame.

[0087] Understandably, this extension allows the prediction method to be directly applied to pipeline systems on moving platforms such as offshore floating LNG production, storage, and offloading (LNG) units. Traditional models for fixed onshore pipelines cannot handle the additional inertial forces introduced by platform motion, leading to prediction failures in marine environments. This invention uses platform motion acceleration as a source term in the physical equations and an input feature of the model, enabling the model to inherently learn and predict the complex coupled dynamics of "platform motion-pipe fluid response." This allows the method to provide early warnings of unique flow risks caused by severe sea conditions, such as slugging and accumulation of liquid within the pipeline during specific roll cycles to form slug flows, or sudden platform acceleration causing a sharp drop in pump inlet pressure leading to cavitation, thereby greatly enhancing the applicability and value of this method in a wider range of more demanding marine engineering fields.

[0088] The workflow of this invention will now be illustrated through two specific embodiments: Example 1 In a 150-kilometer-long onshore LNG pipeline system, connecting an upstream LNG receiving terminal to a downstream city gate station, the pipeline is designed for a pressure of 10 MPa and has a diameter of DN800. Along the route are three intermediate pumping stations, two distribution stations, and multiple shut-off valve chambers. To ensure the safety and efficiency of this long-distance, high-pressure, and cryogenic transportation process, real-time, continuous, and high-precision prediction of the flow state throughout the entire pipeline is required. Traditional methods relying on limited discrete-point monitoring instruments cannot reconstruct complete flow field information within the pipeline and are insufficient for early warning of transient multiphase flow risks caused by rapid valve operation, pump start-up and shutdown, or external heat intrusion, such as slug flow, sudden pressure rise, or localized vaporization. The method and system provided by this invention aim to solve this problem, achieving a leap from "point monitoring" to "field prediction."

[0089] See Figure 1 The overall architecture of this system includes a multi-source heterogeneous data acquisition and preprocessing module, a physical mechanism constraint embedding module, a deep spatiotemporal fusion prediction model module, and a real-time prediction and visualization output module. Each module interacts with a real-time database via high-speed industrial Ethernet, forming a closed loop from data perception to intelligent decision-making.

[0090] First, the multi-source heterogeneous data acquisition and preprocessing module begins operation. (See also...) Figure 3 The core of this module is a distributed sensor network unit deployed at key nodes of the pipeline. In this embodiment, pressure and temperature sensors are installed in pairs on the outer wall of straight pipe sections at 200-meter intervals to monitor the static pressure and temperature of the fluid. High-frequency dynamic pressure sensors and triaxial vibration sensors are added 50 meters before and after the pump station inlet, 1 meter before and after each main valve, and on the outside of all bends greater than 30 degrees to capture local pulsations caused by changes in fluid momentum. Ultrasonic flow meters are used and installed at the outlets of each pump station and the inlets of distribution stations to provide volumetric flow rate data. Acoustic emission sensor arrays are deployed around potentially weak points such as pipe welds and flange connections to collect broadband acoustic signals generated by fluid cavitation, phase change, or micro-leakage. In addition, a distributed fiber optic sensing system is laid in parallel along the entire pipeline. Based on the principles of Raman and Brillouin scattering, it can continuously measure the axial temperature and strain field distribution of the pipeline with a spatial resolution of 1 meter and a temporal resolution of 1 second. All sensors are connected to the regional data acquisition station via an industrial fieldbus, and then aggregated to the data server in the central control room via a fiber optic ring network.

[0091] The data cleaning unit then processes the collected raw sensor data. This unit runs an algorithm based on statistical outlier detection. Its core principle is to take 25 historical data points forward and backward from the current data point for each sensor data stream, forming a sliding window containing 50 neighboring points. The mean and standard deviation of all data within this window are calculated. If the absolute difference between the current data point's value and the mean exceeds 2.0 times the standard deviation, the data point is identified as an outlier. Outliers are removed, and replacement values ​​are generated by an algorithm based on linear interpolation and historical trend extrapolation. For example, if a pressure sensor generates a data point that significantly deviates from the normal fluctuation range due to instantaneous electromagnetic interference, this point will be identified and replaced to ensure the continuity of the data sequence. The system is designed with data quality monitoring and fault tolerance mechanisms. When continuous missing data from some sensors is detected, the system can automatically switch to a lightweight real-time simulator based on physical equations for short-term state extrapolation, or use data from spatially adjacent sensors and the spatiotemporal prediction model itself for interpolation, ensuring continuous and reliable prediction output even when the main model input is incomplete.

[0092] The spatiotemporal alignment unit is responsible for resolving the spatiotemporal inconsistency of data. Based on a high-precision network time protocol, this unit assigns a uniform, millisecond-level timestamp to every data packet uploaded from the distributed sensor network unit. Subsequently, using a 3D geometric model of the pipeline stored in the system—which precisely includes the pipeline's direction, elevation, diameter, material, and the position coordinates of all components—the spatiotemporal alignment unit maps the data from each discrete sensor measurement point to the corresponding spatial coordinates of the 3D model, based on its installation location. For example, a pressure sensor reading installed at KP120+350 is mapped to the corresponding spatial point in the pipeline's 3D model, and the pressure value at that point is used as the representative value for that grid point. For continuous spatial data provided by the distributed fiber optic sensing system, registration is directly performed with the pipeline model according to its inherent spatial resolution. Through this step, all data from diverse sources and with different sampling frequencies are unified under the same time reference and 3D spatial coordinate system, forming a spatiotemporally aligned original data field.

[0093] The feature standardization unit then normalizes the aligned data. This unit calculates the mean and standard deviation of different physical quantities, such as pressure, temperature, flow rate, acoustic emission signal intensity, fiber temperature, and fiber strain, over a long-term historical data period. For the real-time input data stream, the unit performs Z-score standardization, subtracting the historical mean of the physical quantity from the original data value and then dividing by its historical standard deviation. This process transforms all physical quantity data with vastly different dimensions and magnitudes into a standard normal distribution with a mean of 0 and a standard deviation of 1. This eliminates the negative impact of physical dimensions on subsequent neural network model training and accelerates model convergence. Finally, this module outputs a well-organized, time-synchronized, spatially defined, and numerically standardized multidimensional time-series data stream, which serves as the input to the entire prediction system.

[0094] The physical mechanism constraint embedding module constructs and processes in parallel the physical laws describing the fluid motion within the pipe. See also Figure 4 This module is initiated by the governing equations construction unit. Based on the laws of conservation of mass, momentum, and energy, this unit constructs a set of governing equations applicable to the cryogenic, compressible, gas-liquid two-phase transient processes of LNG. The equations consider the compressibility of the fluid, i.e., the change in density with pressure and temperature; they also consider the latent heat absorption and release accompanying the gas-liquid phase transition; and simultaneously, the drag effect of pipe wall friction on the flow is introduced through the Darcy-Weisbach formula. This set of equations is a highly nonlinear partial differential equation set, with variables including fluid velocity, pressure, temperature, gas phase volume fraction, and density.

[0095] The equation discretization unit receives the set of governing equations. To transform them into a computer-processable form and ultimately embed them into a neural network, the unit discretizes the computational domain using the finite volume method. The 150-kilometer-long pipeline is divided into a total of 300,000 hexahedral computational grids in three-dimensional space, with each grid representing a control volume. On each control volume, the governing equations in partial differential form are integrated, and the Gaussian divergence theorem is applied to transform the volume integral into a surface integral, thus obtaining discrete equations describing the changes in physical quantities on each grid. For the time dimension, an implicit time-progression scheme with good numerical stability is used for discretization, with the time step set to 0.1 seconds according to the Courant number condition. After discretization, the original set of partial differential equations is transformed into a large system of nonlinear algebraic equations concerning all state variables at all grid points.

[0096] The loss function construction unit operates based on the discretized results. The core task of this unit is to construct a loss function term that measures whether the neural network's predictions satisfy the aforementioned physical laws. The principle is as follows: for the pipeline's overall state predicted by the neural network at a certain moment, it is substituted into the aforementioned discretized algebraic equations. Since the predicted state is not an exact solution to the equations, the left and right sides of the equations will not be completely equal after substitution, resulting in a residual. This unit calculates the sum of the squares of the residuals of all control equations across all 300,000 computational grids and defines this sum as the physical constraint loss term. Mathematically, the physical constraint loss term is equal to the summation over all computational unit indices from 1 to the total number. The summation term is the square of the L2 norm calculated by the residual operator for the state variables and model parameters of the i-th computational unit. Here, the residual operator represents the discretized control equations, the state variables include the predicted velocity, pressure, temperature, and phase fraction of the grid, and the model parameters are the weights and biases of the neural network. The smaller the value of this loss term, the more the neural network's predictions conform to the physical laws described by the conservation of mass, momentum, and energy.

[0097] The deep spatiotemporal fusion prediction model module is the intelligent core of the system, responsible for performing the actual prediction calculations. (See also...) Figure 2 This module employs an encoder-decoder architecture. The spatiotemporal feature encoder first receives a multidimensional time-series data stream from the preprocessing module. This encoder consists of a gated recurrent unit cascaded with a one-dimensional convolutional neural network. The gated recurrent unit network layer processes the input data sequentially in chronological order. Its internal update and reset gate mechanisms enable it to effectively capture long-range temporal dependencies spanning hundreds of time steps in the data, such as identifying the periodicity of pressure wave propagation or the slow trend of temperature changes. The time-rich feature sequence after processing by the gated recurrent unit is fed into the one-dimensional convolutional neural network layer. These convolutional layers perform sliding convolution operations along the spatial dimension of the pipe axis. Their convolutional kernels can extract spatial local correlation features between data at adjacent locations in the pipe, such as identifying upstream and downstream pressure abrupt changes due to valve throttling or flow pattern changes caused by topographic relief. The encoder output is a feature vector that deeply integrates spatiotemporal information.

[0098] The physical information fusion layer is key to achieving physical consistency in this invention. This layer is located between the encoder and decoder. During model training, this layer passes the predicted states corresponding to the deep features extracted by the encoder to the loss function construction unit of the physical mechanism constraint embedding module, calculating the physical constraint loss term. Subsequently, in the overall objective function of model optimization, in addition to the data loss term measuring the difference between the predicted value and the true label, a weighted physical constraint loss term is forcibly added. The weighting coefficient is an adjustable hyperparameter, determined to be 0.5 in this embodiment through grid search. Its value ranges from 0.1 to 1.0, used to balance the weight between data fitting accuracy and the degree of satisfaction of physical laws. Through the backpropagation algorithm, the model parameters are updated not only in the direction of reducing prediction error but also in the direction of reducing the residuals of the physical equations during optimization. This mechanism uses the partial differential equations describing the nonlinear evolution of LNG multiphase flow as soft constraints, deeply embedding them into the training process of the neural network, thereby driving the model to learn dynamic evolution patterns that conform to physical laws, rather than merely memorizing statistical correlations in the training data.

[0099] The state decoder and predictor receives the feature vectors optimized by the physical information fusion layer. This decoder consists of a fully connected neural network and a deconvolutional network. The fully connected neural network first maps the feature vectors to a high-dimensional latent space. Then, the deconvolutional network begins its work, performing the opposite operation to the one-dimensional convolutional neural network in the encoder. Starting from the compressed feature representation, it gradually restores the size of the feature map through multiple upsampling and deconvolution operations, ultimately outputting a three-dimensional tensor. The three dimensions of this tensor correspond to: the number of grid points along the pipe axis, the time step for future prediction, and the number of channels for the flow state. Specifically, for a prediction of the next 30 minutes (i.e., 180 time steps, each 10 seconds), the decoder outputs a tensor of size "300,000 grid points × 180 time steps × 3 channels". The data in these three channels correspond to the fluid axial velocity value, gas volume fraction value, and static pressure value of each grid point at each future time. Thus, the model completes the mapping and prediction from historical time-series data to the future overall flow state.

[0100] The real-time prediction and visualization output module is responsible for transforming the model's "numerical" predictions into "information" that engineers can intuitively understand. See also... Figure 5This module is first initiated by the result post-processing unit. This unit receives the flow state tensor output by the prediction model. Since the data in this tensor has been standardized, the post-processing unit needs to perform an inverse standardization operation. It calls the historical mean and standard deviation of each physical quantity stored in the feature standardization unit, and performs the inverse operation on each data point of each channel in the tensor, that is, first multiplying by the standard deviation and then adding the mean, thereby restoring the predicted value to engineering data with actual physical units, such as velocity in meters per second, pressure in megapascals, and gas phase volume fraction in percentage.

[0101] The graphics rendering engine then uses this engineering data to drive visualization. The engine incorporates a 3D pipeline model, consistent with the model used by the preprocessing module. The engine retrieves predicted data from the post-processing unit in chronological order. For the velocity field, the engine maps the velocity magnitude of each grid point to a color, rendering a velocity contour map in real-time on the surface or cross-section of the pipeline 3D model. Red represents high-speed zones, and blue represents low-speed zones, dynamically displaying the distribution and evolution of flow velocity along the pipeline. For phase distribution, the engine renders the gas-liquid interface with transparency or a specific color based on the gas volume fraction of each grid point, generating an animation of the phase interface moving within the pipeline, clearly indicating the formation, development, and dissipation of slug flow. For pressure fluctuations, the engine draws pressure contour maps or 3D pressure surfaces, visually displaying the location of high-pressure and low-pressure zones and the strength of pressure gradients. All these graphical elements are displayed panoramically on a large screen in the dispatch center at a refresh rate of at least 10 frames per second.

[0102] The intelligent alarm unit operates synchronously, performing real-time safety analysis on the prediction results. This unit presets a series of dynamic and static thresholds for key parameters. For example, it sets that the instantaneous pressure value at any point in the pipeline must not exceed 1.1 times the design pressure, and the rate of pressure change within 1 second must not exceed 2 MPa per second; it sets a gas blockage risk when the gas volume fraction exceeds 0.8 in a local 1-meter pipe section for more than 10 seconds; and it sets a severe shear flow warning when the velocity difference between adjacent grid points exceeds 5 m / s. The unit continuously scans every grid point and every time step in the prediction data for the next 30 minutes. Once any parameter is detected to exceed the threshold, a tiered alarm is immediately triggered. The alarms are divided into three levels: warning, alarm, and emergency, corresponding to yellow, orange, and red alerts, respectively. Upon alarm triggering, the unit automatically generates a structured report. The report includes the predicted pipeline location of the anomaly, the estimated time of occurrence, the anomaly type, and the severity level. Based on a knowledge base, it provides preliminary handling suggestions, such as "A pressure surge is predicted at KP85+200 in 15 minutes; it is recommended to reduce the output power of upstream pump station 2 in advance" or "A slug flow trend is predicted near KP120; it is recommended to check the air intake in this area." This report is pushed to dispatchers through various means, including audio-visual displays, pop-up windows, and SMS messages.

[0103] The execution flow of the method described in this embodiment strictly follows a predefined sequence of steps. Step S110: Through the aforementioned distributed sensor network unit, raw data from pressure, temperature, flow rate, acoustics, and fiber optic sensing are synchronously acquired at a frequency of up to 100 Hz. This data is then processed continuously by a data cleaning unit, a spatiotemporal alignment unit, and a feature standardization unit to form a regular multi-dimensional time-series input data stream, which is transmitted to the prediction system in real time. Step S120: The deep spatiotemporal fusion prediction model described above is constructed. Its encoder-decoder architecture and physical information fusion layer ensure that the model possesses the ability to learn under physical mechanism constraints. Step S130: Using historical normal operating condition data accumulated over the past year and corresponding flow state label data generated by high-precision simulation software, the model is trained offline. The flow state labels are generated offline using high-fidelity computational fluid dynamics simulation software based on the same pipe geometry and operating parameters. In pipe sections equipped with experimental high-density sensor arrays, data is acquired through direct measurement and data assimilation techniques. Training employed an adaptive moment estimation algorithm optimizer with a batch size of 128, undergoing 500 training epochs until the model's mean absolute percentage error on the independent validation set fell below a preset accuracy threshold of 2.5%. In step S140, real-time acquired and preprocessed multidimensional time-series input data was continuously input into the trained deep spatiotemporal fusion prediction model using a sliding window (30-minute window length, 10-second sliding step). The model performed forward inference computation on a server equipped with a dedicated inference accelerator card, with a single inference time of less than 50 milliseconds. This enabled the model to continuously output predicted tensors of the velocity field, phase distribution, and pressure fluctuations of the entire pipeline within a 30-minute window, with a 10-second cycle. The system's end-to-end prediction latency (from data acquisition to visualization output) was less than 100 milliseconds, meeting the requirements for advanced early warning of rapid transient processes such as slug flow. In step S150, the prediction result tensor is sent to the real-time prediction and visualization output module. After inverse normalization by the result post-processing unit, a dynamic panoramic view is generated by the graphics rendering engine, and the intelligent alarm unit performs real-time scanning and alarm generation based on safety rules, completing the end-to-end process from data to decision support.

[0104] Example 2 This embodiment applies to the internal pipeline system of an offshore floating liquefied natural gas (LNG) production, storage, and offloading (LNG) facility. Unlike long-distance land-based pipelines, this pipeline system is located on a swaying offshore platform. The pipeline is shorter but operates under more complex conditions, involving frequent reliquefaction cycles, mixed transport of LNG from different storage tanks, and the additional inertial forces exerted on the fluid within the pipeline by platform movement caused by sea conditions.

[0105] The system architecture of this embodiment is the same as that of Embodiment 1. See [link / reference]. Figure 1However, specific implementation details have been adapted and enhanced for the marine environment. The distributed sensor network unit of the multi-source heterogeneous data acquisition and preprocessing module, in addition to pressure, temperature, flow, and acoustic emission sensors, adds a high-precision platform motion attitude sensor for real-time acquisition of the platform's roll, pitch, and heave acceleration data. Furthermore, for the high humidity and high salt spray environment at sea, all sensors employ housings and sealing designs meeting the highest protection standards. The data cleaning unit, besides handling outliers caused by electrical noise, is specifically configured with algorithms to identify and compensate for periodic fluctuations in sensor measurements caused by severe platform shaking, separating them from actual fluid fluctuations. The spatiotemporal alignment unit, when mapping data, considers not only the fixed coordinate system of the pipeline but also introduces a dynamic coordinate system determined by the platform's attitude data to accurately reflect the true state of the fluid relative to the moving pipeline.

[0106] The governing equation construction unit of the physical mechanism constraint embedding module explicitly adds a volumetric force source term caused by platform acceleration to the standard mass, momentum, and energy conservation equations. This source term is dynamically calculated based on real-time acquired platform motion acceleration data and is used as an additional term in the momentum equation, thereby incorporating the driving force effect of platform motion on the fluid inside the pipe into the physical constraints. The equation discretization and loss function construction process is similar to that in Example 1, but the governing equations include a time-varying additional source term.

[0107] The architecture of the deep spatiotemporal fusion prediction model module remains unchanged, but its input data includes not only pipeline fluid sensing data but also time-series platform motion data. The spatiotemporal feature encoder needs to learn the coupling relationship between fluid data and motion data simultaneously. The physical constraint loss term in the physical information fusion layer, because the governing equations include platform motion source terms, also forces the model's prediction results to satisfy the fluid motion laws in a swaying coordinate system. This enables the model to predict special offshore conditions such as "whether liquid will accumulate at the low point of the pipeline to form a slug during a specific roll cycle" or "the impact of sudden platform heave on the outlet pressure."

[0108] The graphics rendering engine of the real-time prediction and visualization output module displays the platform's real-time motion attitude curve and key indicator curves of the predicted flow field side-by-side with the standard visualization view, helping operators intuitively understand the causal relationship between motion and flow. The rule base of the intelligent alarm unit has added specific rules for offshore operating conditions. For example, it performs correlation analysis between "platform roll angular velocity exceeding a certain threshold" and "predicted rapid increase in gas phase fraction at pipe bends" to trigger a composite alarm, indicating "motion-induced cavitation risk".

[0109] The execution steps of the method are consistent with those of Example 1, but the data source includes platform motion data, and the historical data used for model training are all from the actual operation and simulation data of the offshore device, ensuring the model's predictive ability for unique offshore conditions. The system makes predictions at a higher frequency (e.g., 1-second cycle) to adapt to faster dynamic changes in offshore conditions.

[0110] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

[0111] In a preferred embodiment, this application also provides an electronic device, the electronic device comprising: The system includes a memory and a processor, the memory storing computer-readable instructions that, when executed by the processor, implement the real-time flow state prediction system for the LNG pipeline transportation system. The computer device can be broadly categorized as a server, terminal, or any other electronic device with the necessary computing and / or processing capabilities. In one embodiment, the computer device may include a processor, memory, network interface, communication interface, etc., connected via a system bus. The processor of the computer device can be used to provide the necessary computing, processing, and / or control capabilities. The memory of the computer device may include a non-volatile storage medium and internal memory. The non-volatile storage medium may store an operating system, computer programs, etc. The internal memory can provide an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface and communication interface of the computer device can be used to connect and communicate with external devices via a network. When the computer program is executed by the processor, it performs the steps of the method of the present invention.

[0112] This invention can be implemented as a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, causes the steps of the methods of embodiments of the invention to be performed. In one embodiment, the computer program is distributed across multiple network-coupled computer devices or processors, such that the computer program is stored, accessed, and executed in a distributed manner by one or more computer devices or processors. A single method step / operation, or two or more method steps / operations, may be executed by a single computer device or processor or by two or more computer devices or processors. One or more method steps / operations may be executed by one or more computer devices or processors, and one or more other method steps / operations may be executed by one or more other computer devices or processors. One or more computer devices or processors may execute a single method step / operation, or execute two or more method steps / operations.

[0113] Those skilled in the art will understand that the method steps of this invention can be performed by a computer program instructing related hardware, such as a computer device or processor, to perform the steps of this invention when executed. Depending on the context, any references herein to memory, storage, databases, or other media may include non-volatile and / or volatile memory. Examples of non-volatile memory include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, magnetic tape, floppy disk, magneto-optical data storage device, optical data storage device, hard disk, solid-state drive, etc. Examples of volatile memory include random access memory (RAM), external cache memory, etc.

[0114] The technical features described above can be combined arbitrarily. Although not all possible combinations of these technical features are described, any combination of these technical features should be considered to be covered by this specification, provided that such combination does not contain contradictions.

[0115] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A real-time prediction system for flow regime in an LNG pipeline transportation system characterized by, The system includes: The multi-source heterogeneous data acquisition and preprocessing module is used to acquire raw sensor data from a distributed sensor network and perform cleaning, spatiotemporal alignment and standardization to generate a multi-dimensional time-series data stream. The physical mechanism constraint embedding module is used to construct a set of control equations describing the multiphase flow process in LNG pipelines and discretize them into physical constraint loss functions that can be embedded in neural network training. The deep spatiotemporal fusion prediction model module is used to receive the multidimensional time series data stream, and perform training and inference based on the physical constraint loss function to generate prediction results of the velocity field, phase distribution and pressure fluctuation of the entire pipeline within the future time window; The real-time prediction and visualization output module is used to perform inverse standardization and visualization rendering on the prediction results and generate alarm information.

2. The LNG pipeline transportation system flow regime real-time prediction system of claim 1, wherein, The multi-source heterogeneous data acquisition and preprocessing module includes: The distributed sensor network unit is deployed at key nodes of the pipeline, including pump station inlets and outlets, valves, bends and straight sections, and includes pressure sensors, temperature sensors, flow meters, acoustic emission sensors and distributed fiber optic sensing systems. The data cleaning unit is configured to use an algorithm based on statistical outlier detection to identify and replace outliers in the original sensor data. The algorithm operates based on a sliding window containing a preset number of neighborhood points and a standard deviation threshold of a preset multiple. The spatiotemporal alignment unit is configured to timestamp all sensor data based on a time synchronization protocol and map the measurement data to a unified spatial coordinate system using a pipeline geometry model. The feature standardization unit is configured to standardize the data processed by the spatiotemporal alignment unit to eliminate differences in physical dimensions and magnitudes.

3. The LNG pipeline transportation system flow regime real-time prediction system of claim 1, wherein, The physical mechanism constraint embedding module includes: The governing equation construction unit is configured to construct a set of governing equations based on the laws of mass conservation, momentum conservation, and energy conservation, applicable to the transient process of cryogenic multiphase flow of LNG, and taking into account fluid compressibility, latent heat of phase change, and pipe wall friction effects. The equation discretization unit is configured to spatially discretize the set of governing equations in the pipeline computational domain using the finite volume method and temporally discretize them using a time-progression scheme to form discretized residual equations. The loss function construction unit is configured to construct a physical constraint loss term from the sum of squares of the discretized residual equations, which is used to measure the deviation between the neural network prediction results and the physical laws described by the set of governing equations.

4. The real-time flow state prediction system for an LNG pipeline transportation system according to claim 3, characterized in that, The control equation construction unit is also used to explicitly add a volumetric force source term caused by the platform motion acceleration to the momentum equation of the control equation set, so as to incorporate the driving force effect of the offshore platform motion on the fluid inside the pipe into the physical constraints; accordingly, the input data of the deep spatiotemporal fusion prediction model module synchronously includes the time series data of the platform motion, and the physical constraint loss term forces the prediction result of the model to satisfy the control equation set including the volumetric force source term.

5. The real-time flow state prediction system for an LNG pipeline transportation system according to claim 1, characterized in that, The deep spatiotemporal fusion prediction model module adopts an encoder-decoder architecture and includes: The spatiotemporal feature encoder is composed of a recurrent neural network for capturing the long-range temporal dependencies of the multidimensional temporal data stream and a one-dimensional convolutional neural network for extracting local spatial features along the pipe axis. The physical information fusion layer, located between the encoder and the decoder, is used during the model training phase to substitute the predicted state corresponding to the deep features extracted by the encoder into the physical constraint loss function, and to add the weighted physical constraint loss term to the model's total loss function to force the model's prediction results to meet physical laws, wherein the weighting coefficient is an adjustable hyperparameter within a preset range. The state decoder and predictor, consisting of a fully connected neural network and a deconvolutional network, is used to decode the feature vectors after fusing physical constraints into flow state tensors covering all computational grid points across the entire pipeline at multiple future time steps.

6. The LNG pipeline transportation system flow regime real-time prediction system of claim 1, wherein, The real-time prediction and visualization output module includes: The result post-processing unit is configured to perform inverse normalization on the flow state tensor output by the deep spatiotemporal fusion prediction model module, and restore it to engineering data with actual physical units. The graphics rendering engine is configured to generate velocity cloud maps, phase interface position animations, and pressure contour maps in real time within the three-dimensional model of the pipeline based on the engineering data. The intelligent alarm unit is configured to perform real-time analysis of the prediction results based on preset safety rules. When the predicted velocity field, pressure gradient, or gas phase volume fraction meets the alarm conditions, it automatically triggers a graded alarm and generates a structured report containing the abnormal location, severity, and handling suggestions.

7. A method for real-time prediction of flow regime in LNG pipeline transportation system characterized by, The method for real-time prediction of flow status in an LNG pipeline transportation system according to any one of claims 1-6 includes the following steps: S1. Through a distributed sensor network deployed in the LNG pipeline system, raw sensor data is collected synchronously, and the raw sensor data is cleaned, spatiotemporally aligned and standardized preprocessed to form multi-dimensional time-series input data. S2. Construct a deep spatiotemporal fusion prediction model with embedded physical mechanism constraints. The model takes the multidimensional time-series input data as input and introduces a physical constraint loss term obtained by discretization of the LNG multiphase flow control equation during the training process to drive the model to learn a dynamic evolution mode that conforms to physical laws. S3. Use historical working condition data to train the deep spatiotemporal fusion prediction model offline to optimize the model parameters; S4. Input the multi-dimensional time-series input data that has been collected and preprocessed in real time into the trained deep spatiotemporal fusion prediction model for forward inference, and output the prediction results of the velocity field, phase distribution and pressure fluctuation of the entire pipeline in the future prediction time domain. S5. Perform inverse standardization and visualization rendering on the prediction results to generate a dynamic panoramic view of the pipeline flow state and perform real-time analysis and alarm.

8. The LNG pipeline transportation system flow regime real-time prediction method of claim 7, wherein, The cleaning process employs an algorithm based on statistical outlier detection. A sliding window containing a preset number of neighboring points is constructed with the current data point as the center. If the absolute difference between the value of the current data point and the mean of the data within the window exceeds a standard deviation threshold based on a preset multiple, it is determined to be an outlier and replaced. The alignment process uses a time synchronization protocol to synchronize timestamps and utilizes a pipeline geometry model to map the measurement data to a unified spatial coordinate system. The standardization process performs standardization calculations on data of different physical quantities to eliminate differences in dimensions and magnitudes.

9. The LNG pipeline transportation system flow regime real-time prediction method of claim 7, wherein, The physical constraint loss term is constructed based on the sum of squares of the residual equations after discretization of the LNG multiphase flow control equations; the total loss function of the deep spatiotemporal fusion prediction model is the sum of the data loss term and the weighted physical constraint loss term, wherein the coefficients used for weighting are adjustable hyperparameters within a preset range. By optimizing the total loss function, the prediction results of the model can simultaneously satisfy the data fitting accuracy and physical law constraints.

10. The LNG pipeline transportation system flow regime real-time prediction method of claim 7, wherein, The distributed sensor network additionally includes motion attitude sensors for collecting platform motion acceleration data; in the step of constructing a deep spatiotemporal fusion prediction model, the LNG multiphase flow control equation includes a volume force source term dynamically calculated from real-time platform motion acceleration data, and the multidimensional time-series input data synchronously includes time-series data of platform motion, so that the trained model can predict the coupling effect between platform motion and fluid flow in the pipe.

Images