Runtime-constrained day-ahead and real-time electricity price forecasting in PJM’s BGE zone: stacked ensembles and residual diffusion
TL;DR
A runtime-constrained framework for day-ahead (DA) and real-time (RT) locational marginal price forecasting in PJM’s Baltimore Gas and Electric zone is developed and expected demand is identified as the dominant predictive input, with stronger responses during peak and cold conditions.
Abstract
Accurate short-term electricity price forecasting requires evaluation using only information available before market submission. This study develops a runtime-constrained framework for day-ahead (DA) and real-time (RT) locational marginal price forecasting in PJM’s Baltimore Gas and Electric zone. Here, runtime-constrained refers to information availability at the forecasting cutoff, not computational execution time. Forecasts are issued at 10:00 AM Eastern Prevailing Time, one hour before the 11:00 AM DA deadline. Strict RT predictors use completed observations through 09:00 and only RT labels available by the forecasting runtime. The stacked ensemble achieves median DA and strict RT mean absolute errors (MAE) of 8.26 and 14.60 USD MWh−1, respectively, with corresponding root mean squared errors of 10.50 and 21.77 USD MWh−1. Relative to a runtime-safe 48 h persistence benchmark, MAE is reduced by 29.6% for DA and 20.1% for strict RT. Extra Trees and random forest remain statistically competitive with stacking, so stacking is treated as a strong operational backbone rather than a universally dominant model. Rolling conformal intervals achieve pooled 90% coverage of 88.8% for DA and 88.2% for strict RT, while raw residual diffusion is less well calibrated. A separate auxiliary reliability-constrained effective-information adaptive memory (RC-EIAM) analysis adaptively recalibrates residual-diffusion forecasts using previously verified lead-specific information and then applies a reliability constraint to the prediction interval. Applying RC-EIAM to strict-RT diffusion increased pooled 90% coverage from 77.8% to 89.9% and improved the continuous ranked probability score by 4.9%. It also reduced interval score by 10.2%, although this improvement was descriptive and not statistically resolved (HAC p = 0.188). The improved coverage required wider intervals. DA SHapley Additive explanation (SHAP) and load-forecast sensitivity analyses identify expected demand as the dominant predictive input, with stronger responses during peak and cold conditions.