DEVELOPMENT OF A DECISION SUPPORT SUBSYSTEM FOR PREDICTIVE THERMAL CONTROL IN PIPELINE SYSTEMS

Abstract

The developed Decision Support System (DSS) for predictive thermal control of pipelines is based on a standardized digital twin that represents the geometry, materials, environmental conditions, and sensor layout of the pipeline. This virtual model enables accurate simulation of heat transfer, detection of anomalies, modeling of heat losses, and generation of control actions in real time. At the core of the system lies a forecasting module that predicts future temperature distributions based on current conditions and physical heat transfer patterns. These predictions are generated using empirical models or neural network-based predictors, trained to high accuracy and continuously updated as new data becomes available. The DSS incorporates fuzzy logic to manage uncertainty by using linguistic variables such as “overheating”, “low gradient”, and “instability”. Sensor inputs are transformed into degrees of membership to these concepts, which are then processed through expert-defined fuzzy rules. The outcome is translated into concrete control actions, such as adjusting heating power, through a defuzzification process. A reinforcement learning module is also integrated to autonomously refine control strategies. By interacting with the pipeline model, it learns optimal decisions that maintain thermal stability, minimize energy consumption, and prevent hazardous conditions. To capture causal dependencies between pipeline segments, the DSS employs Bayesian networks, which model how thermal effects propagate along the structure. These probabilistic graphs help assess how local heating can influence adjacent segments and overall system behavior. In experimental validation using M2PLink simulation software, the DSS demonstrated high predictive accuracy. The comparison between predicted and simulated temperature distributions showed strong alignment, confirming the system’s ability to replicate real thermal dynamics and respond effectively to operational changes. It is substantiated that the system provides the ability to scale to complex network pipeline structures, including branched trunks and multi-component nodes. The extended integration of the digital twin with real sensor data allows to detect potential deviations and predict the development of faults long before their actual manifestation. It is established that the improved forecasting process contributes to adaptive control of thermal regimes, which reduces operating costs and increases the overall reliability of the pipeline infrastructure. It is proven that the use of a combined approach, which combines methods of artificial intelligence, mathematical modeling and fuzzy logic, increases the system’s resistance to noise, incomplete data and nonlinear behavior of physical processes. An additional pilot experimental evaluation was also conducted using an independent SCADA dataset for monitoring oil and gas pipeline systems, covering 12 months of continuous measurements sampled at a frequency of 1 Hz. Temperature forecasting was implemented using a feedforward neural network focused on short-term incremental temperature changes, combined with control based on the Proximal Policy Optimization algorithm. The obtained results confirmed high forecasting accuracy and effective adaptive control: the average duration of the complete decision-making cycle did not exceed 18.5 ms, while the cumulative discounted reward increased by 74% with a simultaneous reduction of its variance by more than 60%.