Grid operators managing high renewable energy penetration face a critical challenge: predicting net electricity demand (total load minus renewable generation) accurately over extended periods. Traditional forecasting methods struggle with the quasi-periodic nature of wind and solar output combined with unpredictable weather-driven fluctuations. HarmoNet addresses this problem through an innovative dual-domain architecture that captures both slow, predictable trends and rapid, short-term variations in power flows.
The model employs a hybrid approach combining convolutional neural networks for local pattern recognition with transformer layers for capturing long-range dependencies across the power system. By encoding coupled low-frequency and high-frequency signal components with multi-scale temporal patterns, HarmoNet effectively decomposes complex net load dynamics into manageable components. A quantile regression framework further estimates confidence intervals around point forecasts, providing grid operators with probabilistic bounds essential for reliable reserve scheduling.
Testing on real-world hourly data from Belgium, Bulgaria, and Italy spanning four years (2016–2019) demonstrated substantial improvements. The model reduced mean absolute error by 14.2% on average, with peak reductions exceeding 22.5% for 30-day horizons. Performance gains were particularly pronounced during high-volatility periods, peak demand spikes, and steep ramps—exactly the stress conditions where grid stability is most threatened.
Uncertainty quantification proved equally important. HarmoNet reduced the Winkler score (a standard probabilistic accuracy metric) by 28.4% and pinball loss by 13.0%, meaning operators receive tighter, more reliable confidence intervals for operational decision-making. These metrics directly translate to better reserve scheduling, reduced emergency curtailment of renewable generation, and lower overall system costs.
The research draws data from the Open Power System Data platform using eight meteorological covariates as inputs. Successful deployment of such models could enable European grid operators to integrate substantially higher renewable capacity while maintaining reliability, supporting the region's ambitious net-zero targets.
