کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

An Uncertainty-Gated Multi Layer Perceptron for Decomposed Probabilistic Net Load Forecasting

نویسندگان :

Sina Hossein Beigi Fard¹ Amir Hossein Baharvand² Meysam Doostizadeh³

1- دانشگاه لرستان 2- دانشگاه لرستان 3- دانشگاه لرستان

کلمات کلیدی :

short-term net load forecasting،probabilistic forecasting،quantile regression،multi-layer perceptron،uncertainty quantification

چکیده :

Short-term net load forecasting is essential for secure and economical grid operation under increasing penetrations of variable renewable generation. Beyond point accuracy, operators require reliable uncertainty estimates to support reserve sizing and risk-aware scheduling. Existing approaches either focus on deterministic forecasts or provide probabilistic outputs that are difficult to interpret when Net Load (NL) is formed by combining multiple uncertain components, especially under time-varying uncertainty driven by weather effects. Therefore, this work aims to deliver a practical probabilistic net load forecasting framework that produces calibrated prediction intervals while maintaining source-level traceability and feature-level interpretability. The NL is formulated to learn component-wise probabilistic forecasts for load, photovoltaic, and wind using an uncertainty-gated multi-layer perceptron that outputs monotone quantiles via a softplus-based ordered head. On the evaluated dataset, the proposed model achieves the lowest point-error among compared baselines while providing reliable uncertainty with 90% interval coverage. Explainability results indicate that temporal encodings and key weather data dominate the median prediction and uncertainty width, offering consistent, physically plausible insights. The proposed framework provides an operationally meaningful balance between accuracy, reliability, and interpretability, enabling uncertainty-aware decision-making for scheduling and reserve procurement in renewable-rich power systems.