GAPLO: Self Evolution Swarm Agents Framework for Cross Border Pipelines Operational Optimisation with Implementation in China's Energy Supply

Abstract China's Belt&Road long-distance oil&gas transition pipelines (CLGP) play a crucial role in ensuring China's oil and gas (O&G) supply. According to multiple properties like long distances, cross countries, and multiple pipelines synergy, these pipelines are influenced by multiple issues including climatic, geological, geopolitical, multinational workforce gaps, commodity market dynamics, and multi-lateral energy consumption. High decision precision is required in formulating O&G transmission policies. Traditional scheduling policies primarily rely on the engineers’ experiences, coordinate through telephone calls or by scheduling meetings, however, owing to language barriers, inefficiencies in scheduling processes, and the inability to simulate uncertainties associated with unforeseen events, limitations still remained regarding to efficiency of scheduling for variety of non-planned incidents, this remains significant room for continuous improvement in energy supply assurance plans and overall efficiency. The manuscript proposes a self-evolution swarm agents framework (GAPLO) that enables the understanding of CLGP scheduling decision. Inspired by the feedforward neural architecture, GAPLO introduced an agentic neural based approach, to achieve the self-refinery performance in terms of workflows generation, scheduling events simulation, and decision making recommendations. The novel architecture enables autonomy generate/reduce multi-roles agents in response to CLGP operation events and make decisions. By incorporating the mixture-of-experts (MOE) architecture, the framework enabling the simulation and interpretation of multiple pipelines operational policies. To ensure accurate decision generation, and facilitate autonomous learning and optimisation of CLGP operational strategies, GAPLO adopts a dual feedback optimisation mechanism combining agents and human-in-the-loop (HIL). Utilising scheduling, inventory, weather, geopolitical, shipping information, terminal consumption, and financial data from 6 CLGP pipelines spanning over 11,000 km, covering the China-Northeastern/Southern/Central Asia regions, to validate the feasibility of autonomous decision support for the CLGP system based on GAPLO assisted. By employing scheduling data from 6 CLGP over the period from 2020 to 2024, and integrating it with 30 instances of scheduling events with corresponding responses extracted from an online knowledge-base as validation materials, the results demonstrate that the integration of GAPLO with deep-thinking, dual feedback based HIL interaction and MOE can effectively support autonomous multiple CLGP scheduling, especially under abnormal events triggered situations. With subjective and objective accuracy distributed within 77%~95+%, this means that GAPLO could be an innovative evolutionary swarm agent framework for auonomy construction of generative agent clusters for multi-objectives CLGP decision supports, with actual use cases designed to validate the feasibility of whole scheduling procedures from issue spotting to scheduling strategies creation among the lean operation of long distance cross-countries pipes.