Point Forecasts to Probability Clouds — Probabilistic Electricity Price Forecasting

Generative probabilistic forecasting of day-ahead electricity prices with calibrated uncertainty

Overview

I built a probabilistic forecasting pipeline that turns single-point predictions of day-ahead electricity prices into scenario clouds with calibrated uncertainty. The model—UWIAE-GPF (Univariate Weak-Innovation Autoencoder for Generative Probabilistic Forecasting)—learns spike-and-tail behavior typical in electricity markets and produces both point trajectories and risk bands (90/50/10% intervals).

Duration: Jan 2025 – May 2025
Role: Solo Developer Institution: Cornell University Supervisor: Professor Lang Tong

Why it matters: Operators and traders need not only accurate forecasts but also confidence bounds to plan for best/worst cases.

Data & Task

  • Market: NYISO Day-Ahead LBMP (N.Y.C. zone), 2018–2023, hourly.
  • Setup: Rolling windows of past 1–28 days → forecast next-day (24 h) price path.
  • Split: Train (2018–2022), Test (2023).
  • Eval: Normalized errors (e.g., NMSE) + coverage of prediction intervals.

Method

  • UWIAE-GPF: Innovation-autoencoder with reconstruction/innovation critics + adversarial generator to model price dynamics and rare spikes.
  • Deterministic baselines: PatchTST, iTransformer, Informer, GRU, NLinear, StemGNN.
  • Probabilistic baselines: TimeGrad (diffusion), RealNVP/MAF flows (GRU/Transformer-conditioned).
  • Uncertainty: After convergence, draw 1,000 latent samples per window to form scenario ensembles → mean/median paths + central bands (90/50/10%).

Results

  • Best NMSE = 0.093, outperforming deterministic Transformers and earlier generative baselines.
  • ↓65.3% NMSE vs PatchTST (best deterministic) and ↓39.2% vs original WIAE, with ~4× faster per-epoch training than PatchTST.
  • Interval bands capture spikes better, improving risk-aware planning.

Implementation

  • Stack: Python, PyTorch, NumPy, Pandas, matplotlib; Linux; experiment tracking via scripts.
  • Deployment angle: Saves ensembles & summary stats for downstream dashboards or bidding logic.

Highlights

  • Probabilistic wins → lower risk: Calibrated intervals and diverse scenario paths.
  • Scalable: Trains efficiently while modeling heavy tails/spikes.
  • Transferable: Template can be adapted to other volatile time series.

Artifacts

  • Poster (PDF): Download
  • Contact: haixizhang02@gmail.com