Counterfactual load forecasting with LLM-structured events and representation learning

News-reported social, environmental, and grid events can substantially reshape electricity demand, yet quantifying how such events perturb forecasted load trajectories remains largely unaddressed. Existing news-augmented forecasting studies use news as auxiliary features to reduce prediction error but cannot answer counterfactual questions about demand under alternative event conditions. Counterfactual forecasting naturally formulates this problem by comparing the factual trajectory under observed news with counterfactual trajectories under alternative treatments, such as removing events or injecting hypothetical scenarios. However, counterfactual outcomes are inherently unobservable, so the prediction error cannot be directly minimized from data. This challenge is compounded by confounding: news occurrence is entangled with weather, calendar, and historical load conditions that independently affect demand. Guided by a generalization bound for continuous treatments that decomposes the unobservable counterfactual error into weighted factual loss and representation balance, this paper proposes the News-Aware Counterfactual Load Analysis Framework (NACF) to control this error upper bound through observable training objectives. NACF converts unstructured news into structured event streams via an offline large language model, encodes these as continuous semantic treatments with a no-news baseline, and estimates treatment-dependent load trajectories through a varying-coefficient response network with learned sample reweighting and Integral Probability Metric (IPM)-based representation balance regularization. Experiments on an Australian electricity-demand dataset show that NACF remains competitive in factual forecasting while providing evidence of treatment-intensity structure, improved representation balance, and interpretable demand perturbations in synthetic interventions and real event case studies.