Since there are still sufficient and affordable conventional energy in the next three decades, in which the energy riding of fossil fuels exceeds a fully sustainable scene with zero emission scene, the oil and gas industry is aimed at more economically demanding fields in which production systems become lower and complex (Nasser, 2022). The ultimate goal of the oil and gas industry is to identify the optimal field development plan in order to create carbonate accumulation, at the same time to minimize the costs and maximize the income. Therefore, it creates a persistent need to develop tailor -made and economically efficient systems for surface network models that play an important role in minimizing investment expenditure (He et al., 2019; Liu et al., 2019).
When optimizing surface networks, the pressure drop is a key parameter when identifying the optimal pipeline layout and maintaining reservoir productivity (DBouk et al., 2021a, 2021b). The lower the pressure drop, the higher the fountain productivity. In contrast, a high -pressure waste system meets significant problems with flow protection, which leads to high well -headed pressure boundaries and thus to a reduced fountain productivity.
In addition to topological complexity, surface networks encounter a multi -phase current with the multiphase current with the marketing of different liquids from several fountains, zones and fields. Therefore, the pressure drop for the pressure in such systems is an extremely complicated problem from the perspective of fluid mechanics, since this large production system has extreme pressure and temperature changes, which leads to complex phase changes in the entire system (YAQUB et al., 2020).
Correlations are typically used to estimate the pressure drop in a pipeline. These correlations depend on pipeline -specific parameters such as diameter, wall thickness, roughness, length and inclination. The correlations also depend on multiphasic flow-specific parameters, including hydrocarbon liquid rate (oil rate), gas oil ratio (GOR), water-fluid ratio (water neckline, WCT), phase visity, liquid temperature and pipe line output pressure. In view of the large number of parameters and practical and cost restrictions, correlations for different sets of different parameters and for different parameter areas must be coordinated (CHAARI et al., 2020). Therefore, practically each pipe can use different correlations for a production network of pipes. Even for the same pipe, the correlations may have to change if the flow conditions change due to the significant increase in gor and/or WCT and the corresponding pressure conditions.
Artificial intelligence techniques for multiphase flow problems show an improved performance compared to conventional methods (Al Wahaibi et al., 2024). Although the recent studies showed ML's ability to predict pressure drop or pressure gradients by pipelines, the ML models are based on limited data set size and limited feature areas (Al Wahaibi et al., 2024; Hafsa et al., 2024). Therefore, the models may need to be retraining in order to absorb the areas that can occur in real production networks. In addition, data records are used that are accessed either from the literature as experimental data or from narrow field data. A further restriction of the previous studies is that some field parameters, if used under real conditions, have to be changed to meet the ML models inputs. Therefore, there is a modular approach to the development of ML models that predict the pressure drop by piping under any configuration and at the same time treat oil and gas field applications.
In this article, we deal with the restrictions associated with the use of correlations and existing ML models to predict the pressure drop in pipelines by drastically improving the experimental data record and field data with synthetic data records in order to address the inherent challenge of limited data records. In view of the scarce test and field data, the training of a reliable ML model is a real challenge, which is therefore addressed in this article by using synthetic data instead of using experimental and field data. We use conventional empirical correlations and mechanistic models to create large data records that cover the comprehensive areas of different physical parameters. The measured data are typically sparse, but important to calibrate correlations and models. Measured experimental and field data are not only used for training the ML models. They are essentially used to train empirical correlations or mechanistic models that are used to produce reliable synthetic data. As a result, our approach uses indirectly or even directly available real and experimental data by significantly increasing data with synthetic data that is generated from models, which in turn are calibrated with the existing, typical sparse data.
In a typical workflow, various correlations are coordinated using the available experimental and field data for various parameter sets and/or parameter areas. We try the multi-dimensional parameter space and use various correlations to create a synthetic data set that covers the multi-dimensional parameter space and trains an ML model using this synthetic data set. The important aspect of the synthetic data set is the fact that it can be increased with new data if an unexpected set of parameters must be taken into account. This is followed by a new, generic ML model. Therefore, this synthetic data record does not have to be re -generated for each field and in any case. It can be used in an generic manner and can be updated and incremented in order to become increasingly comprehensive and more generally.
Our approach to the development of ML models for the pressure waste forecast by pipelines using synthetic data is modular. It enables the creation of ML models under any configuration while treating real applications and the underlying physics. The input functions for our models are selected as frequent parameters in the field, since it is worth having a prediction tool that can estimate the pressure drop based on the available data and without changes. The approach can be applied to remedy a complexity area: from simple individual phase, individual components, isothermal flow to the multi-phase flow fluids under non-isotherms. Depending on the complexity of the problem, the approach can use standard -ML algorithms and hybrid methods. Therefore, our modular approach is not limited to areas, features or even ML algorithm, which leads to representative, robust and efficient ML models that can be used in real field scenarios.
A central result of this work is the development and integration of these ML models into a conventional network kiln as an alternative to correlations or mechanistic models. This is the first fully ml-based network clasp, which evaluates the pressure drop by networks of multiphase pipeline networks.
An additional innovation in this article is the concept of the “Network Upscaling”, ie the automated and systematic process, which reduces the number of segments in the network model without significantly influencing the results. In view of the complex geometry of real networks, which leads to a large number of segments to present them, the network -upscaling reduces the size of such networks and thus reduces the processing time drastically to maintain the integrity of the results. This is particularly important if the network solution is repeatedly estimated. A large number of wells, forbidden areas and topological complexities. (Ghorayeb et al., 2023).
- •
Scalable synthetic data generation das proposed framework uses calibrated empirical correlations and mechanistic models to create large, high-quality synthetic data sets. It enables incremental updates of the data record to cover new parameter areas, which requires the need to regenerate data sets for each new field or scenario, and therefore enables the training of robust ML models despite limited experimental and field data.
- •
Modular ML Framework-and Workflow enables ML models to be created under various river configurations (e.g. single/multiphase, isothermal/non-isotherm). It supports a wide range of ML algorithms and adapts to different complexity levels and only uses frequently available field parameters as input functions to ensure practical applicability without requiring additional setup.
- •
Full of ML-based network Solver-We replace conventional correlations and mechanistic models within a pipeline networks with a generic, updated ML model for the pressure waste forecast.
- •
Network upscaling technology-like reduces the number of pipeline segments in complex networks without compromise accuracy and significantly improves arithmetic efficiency in large-scale field optimization workflows.
In the following we describe relatives work and discuss associated restrictions. Then we give an overview of the methodology, followed by a detailed description of your steps. Then we discuss the production of the ML model used in the ML-based network climb. Finally, we compare the results of the ML-based networks that we have developed with those received by a commercial simulator.