Market-Informed Multi-Step Forecasting of Intraday Net Load with Operational Risk Screening

The increasing penetration of wind and photovoltaic generation has intensified net-load variability and increased the difficulty of intraday forecasting and operational risk identification. This paper proposes a market-informed multi-step net-load forecasting and risk-screening framework based on Chronos2-Small. Low-rank adaptation (LoRA) is employed for parameter-efficient fine-tuning, while historical net load, meteorological variables, system operating conditions, day-ahead prices, historical real-time prices, and temporal features are jointly incorporated. The model outputs P10, P50, and P90 quantiles for 4, 8, and 12 h forecasting horizons. Experiments on 8544 15 min observations from the Jiangsu power system show that the proposed model achieves nMAE values of 2.72%, 4.22%, and 4.45% for the three horizons, respectively. Across three seeds, MAE is reduced by 37.23%, 33.02%, and 35.46% relative to covariate-extended PatchTST. SHAP analysis shows that the relative importance of day-ahead market signals among the five exogenous groups increases from 32.01% to 65.21% as the forecasting horizon extends. A three-level operational risk mechanism is further constructed using the relative P10–P90 interval width and P50 net-load ramp, with day-ahead prices providing an independent market-sensitive label. The Watch and High states capture 95.41–97.58% of independently defined large-ramp events, with a precision of 23.69–26.41%. Pinball loss, interval score, coverage, and width are used to assess probabilistic performance.