Auto-Regressive HMM Inference with Incomplete Data for Short-Horizon Wind Forecasting

Chris Barber, Joseph Bockhorst, Paul RoebberNeurIPS 2010

Introduction

Integrating wind energy into the electrical grid is a delicate balancing act. Because cost-effective, large-scale energy storage remains elusive, power production must match consumption in real-time. Accurate Short-Term Wind Forecasting (STWF)—predicting wind behavior 0.5 to 6 hours in advance—is worth billions to the energy sector, yet it remains notoriously difficult.

This paper introduces a "regime-aware" approach to STWF using Auto-Regressive Hidden Markov Models (AR-HMMs). By developing a novel approximate inference method, the authors successfully handle the messy reality of missing sensor data, outperforming standard baselines and uncovering meteorologically meaningful weather patterns.

The Challenges of Short-Term Wind Forecasting

The authors identify two primary hurdles in STWF:

  1. Weather Regimes: Wind behavior is not static. A "regime shift," such as a passing weather front, can completely alter how wind variables interact across different geographic locations.
  2. Incomplete Data: Real-world sensor networks are inherently flaky. Observations are frequently missing, delayed, or corrupted. While Hidden Markov Models (HMMs) are excellent at representing shifting regimes, performing exact inference on them when data is missing is an NP-hard problem.

Modeling Regimes with AR-HMMs

To address these challenges, the authors propose using Switching Conditional Linear Gaussian (CLG) models, specifically Auto-Regressive Hidden Markov Models (AR-HMMs). In this architecture, a discrete hidden variable represents the current "weather regime." This regime dictates the linear relationship between current and past wind speeds.

Graphical models illustrating the progression from simple Markov chains to the proposed spatial-temporal AR-HMM with missing data.

To solve the NP-hard inference problem caused by missing data, the authors introduce a pruning-based approximate inference method. Because weather regimes typically persist for several hours, the probability space is dominated by just a few likely sequences. They tested three pruning strategies to make the model computationally tractable:

  • PRIOR(n): Selecting the $n$ most likely regime sequences based on prior transition probabilities.
  • HOMO: A simplified homogeneous version assuming no regime changes occur within a small time window.
  • POST(n): Selecting sequences based on a fast approximation of posterior likelihood using local evidence.

By focusing only on these high-probability "paths," the model can efficiently process incomplete observations without sacrificing predictive power.

Datasets and the "Gapiness" Problem

The study utilized four years of publicly available hourly surface wind data (2006–2010) from the National Climatic Data Center (NCDC). The evaluation focused on two geographically distinct regions in the U.S.: Wisconsin (WI) with 21 observation sites, and the Pacific Northwest (PNW) with 24 observation sites.

A defining characteristic of these real-world datasets is their "gapiness." Only about 6% of the recorded hours contained complete data across all sites. This severe sparsity highlights the absolute necessity of a model designed to handle missing values gracefully.

Key Findings and Performance

  • Beating the Baseline: In very short-term forecasting (under 2 hours), "persistence"—the assumption that the future will look exactly like the present—is a surprisingly tough baseline to beat. Most physical weather models actually perform worse than persistence at these horizons. However, the single-regime AR-LG(1) model significantly outperformed persistence forecasts for the 1-to-2-hour horizon in both regions.
  • Multi-Regime Gains: Moving from a single-regime model to a multi-regime model (AR-LG(2) and above) resulted in a massive jump in log-likelihood. This indicates that the AR-HMM was vastly superior at capturing the underlying probability distribution of dynamic wind behavior.
  • The Value of Regime Diversity: The "HOMO" approach performed surprisingly well. The authors suggest this is because assuming a single regime persists provides a diverse set of "expert" opinions, which is more robust for prediction than selecting several very similar sequences that only differ by a single-hour transition.
  • Spatial Intelligence: The model successfully leveraged "off-site" evidence. By analyzing data from neighboring stations, it could make better local predictions, even when the target station had missing data.
  • Meteorological Grounding: The model did not just optimize abstract metrics; it "discovered" real weather. Visualizations of the learned regimes corresponded to clear diurnal (day/night) cycles and seasonal patterns, proving the AR-HMM successfully captured physical reality.

Limitations and Future Directions

The authors acknowledge a few limitations. First, the study used data from sensors 10 meters above ground, whereas commercial wind turbines operate at roughly 80 meters. While the mathematical methods should transfer, this was a necessary compromise based on NCDC data availability. Second, at longer horizons (5–6 hours), the model's performance in Wisconsin began to degrade relative to persistence, suggesting that "unrolling" the auto-regressive model over many steps can lead to accumulating errors.

Looking forward, the authors suggest their approximate inference method has broad applicability. It could be deployed in other domains beyond meteorology—such as epidemiology or econometrics—where switching linear models must contend with the reality of missing data.

Summary updated

BibTeX
@inproceedings{barber2010auto,
  author = {Chris Barber and Joseph Bockhorst and Paul Roebber},
  title = {{Auto-Regressive HMM Inference with Incomplete Data for Short-Horizon Wind Forecasting}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2010},
}
Figure