Large language model for uncertainty formalisation and decision support in coupled EV charging and ICT workloads

Abstract Urban distribution systems may supply electric vehicle (EV) charging facilities and grid‐connected computing facilities through the same distribution feeder or transformer, creating a coupled flexibility‐allocation problem when mobility‐energy and computing‐service demands become simultaneously stressed. This paper addresses the allocation of limited feeder‐side electrical flexibility between EV charging and information and communication technology computing workloads so as to reduce operating cost and tail‐risk exposure while evaluating service trade‐offs and network‐feasibility outcomes. The proposed large language model (LLM)‐enabled decision‐support framework establishes a structured uncertainty‐to‐decision pipeline comprising domain‐guided tokenisation, context‐aware semantic interpretation, and deterministic decision compilation. Historical uncertainty is incorporated into conditional‐value‐at‐risk (CVaR)‐based day‐ahead envelope planning, while a low‐rank‐adaptation‐enabled LLM organises real‐time numerical, symbolic, and cross‐zone evidence into schema‐constrained operating states, risk levels, flexibility descriptors, and action priorities. These semantic outputs are converted into candidate actions assessed through reserve, service, zonal‐boundary, workload‐offloading, and alternating‐current power‐flow checks; unresolved failures are recorded rather than certified as feasible actions. On the IEEE 33‐bus/four‐zone system, the proposed method reduces mean total cost and CVaR by 16.90% and 9.34%, respectively, relative to rule‐based‐compilation. On the MATPOWER case141/eight‐zone system, the 95th‐percentile runtime is 37.16 s, supporting computational tractability at the evaluated scale.