Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SARIMAX-Based Power Outage Prediction During Extreme Weather Events

Haoran Ye, Qiuzhuang Sun, Yang Yang

TL;DR

This work presents a SARIMAX-based framework for short-term power outage forecasting during extreme weather, leveraging a two-stage feature extraction pipeline to reduce 84 weather features to 37 representative inputs augmented by temporal embeddings and multi-scale lag features. Each of 83 counties is modeled with an independent SARIMAX$(1,0,1)$ using exogenous weather inputs, with a robust training pipeline that includes constant/correlation preprocessing and hierarchical fitting with fallbacks to ARIMA or historical means. The approach achieves $RMSE=177.2$, an $8.4\%$ improvement over a baseline of $RMSE=193.4$, demonstrating the value of careful feature engineering and robust optimization for resilience forecasting. The study also reveals that simpler statistical models can outperform complex neural architectures in data-limited, non-stationary outage scenarios, and discusses future directions including spatiotemporal models and rare-event techniques to further enhance predictive performance.

Abstract

This study develops a SARIMAX-based prediction system for short-term power outage forecasting during extreme weather events. Using hourly data from Michigan counties with outage counts and comprehensive weather features, we implement a systematic two-stage feature engineering pipeline: data cleaning to remove zero-variance and unknown features, followed by correlation-based filtering to eliminate highly correlated predictors. The selected features are augmented with temporal embeddings, multi-scale lag features, and weather variables with their corresponding lags as exogenous inputs to the SARIMAX model. To address data irregularity and numerical instability, we apply standardization and implement a hierarchical fitting strategy with sequential optimization methods, automatic downgrading to ARIMA when convergence fails, and historical mean-based fallback predictions as a final safeguard. The model is optimized separately for short-term (24 hours) and medium-term (48 hours) forecast horizons using RMSE as the evaluation metric. Our approach achieves an RMSE of 177.2, representing an 8.4\% improvement over the baseline method (RMSE = 193.4), thereby validating the effectiveness of our feature engineering and robust optimization strategy for extreme weather-related outage prediction.

SARIMAX-Based Power Outage Prediction During Extreme Weather Events

TL;DR

This work presents a SARIMAX-based framework for short-term power outage forecasting during extreme weather, leveraging a two-stage feature extraction pipeline to reduce 84 weather features to 37 representative inputs augmented by temporal embeddings and multi-scale lag features. Each of 83 counties is modeled with an independent SARIMAX using exogenous weather inputs, with a robust training pipeline that includes constant/correlation preprocessing and hierarchical fitting with fallbacks to ARIMA or historical means. The approach achieves , an improvement over a baseline of , demonstrating the value of careful feature engineering and robust optimization for resilience forecasting. The study also reveals that simpler statistical models can outperform complex neural architectures in data-limited, non-stationary outage scenarios, and discusses future directions including spatiotemporal models and rare-event techniques to further enhance predictive performance.

Abstract

This study develops a SARIMAX-based prediction system for short-term power outage forecasting during extreme weather events. Using hourly data from Michigan counties with outage counts and comprehensive weather features, we implement a systematic two-stage feature engineering pipeline: data cleaning to remove zero-variance and unknown features, followed by correlation-based filtering to eliminate highly correlated predictors. The selected features are augmented with temporal embeddings, multi-scale lag features, and weather variables with their corresponding lags as exogenous inputs to the SARIMAX model. To address data irregularity and numerical instability, we apply standardization and implement a hierarchical fitting strategy with sequential optimization methods, automatic downgrading to ARIMA when convergence fails, and historical mean-based fallback predictions as a final safeguard. The model is optimized separately for short-term (24 hours) and medium-term (48 hours) forecast horizons using RMSE as the evaluation metric. Our approach achieves an RMSE of 177.2, representing an 8.4\% improvement over the baseline method (RMSE = 193.4), thereby validating the effectiveness of our feature engineering and robust optimization strategy for extreme weather-related outage prediction.

Paper Structure

This paper contains 27 sections, 5 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Two-stage Feature Extraction
  4. Stage 1: Data Quality Enhancement
  5. Stage 2: Feature Engineering
  6. Data Augmentation
  7. Temporal Embeddings
  8. Lagged Features
  9. Lagged Outages:
  10. Lagged Tracked Indicator:
  11. Lagged Weather Variables:
  12. Model Training
  13. SARIMAX Model Specification
  14. Training Pipeline
  15. 1. Feature Preprocessing.
  16. ...and 12 more sections

Figures (3)

  • Figure 1: The workflow of the proposed SARIMAX-based prediction model.
  • Figure 2: Comprehensive PCA visualization for Michigan weather features (April--July 2023). (a) 3D scatter (PC1--PC2--PC3) showing temporal dispersion of statewide atmospheric states. (b) Loading heatmap: contribution of original features to the leading PCs. (c) Explained variance ratio and cumulative curve (dashed line marks the 80% threshold; first 20 PCs reach 80.51%). (d) Distribution ranges of PC1/PC2/PC3 scores, indicating spread and potential regime transitions. (e) Bar comparison of top absolute loadings for PC1--PC3 supporting semantic labeling: PC1: temperature--radiative stability pattern; PC2: atmospheric moisture and stability suppression pattern; PC3: surface energy flux and boundary-layer turbulence pattern.
  • Figure 3: Temporal evolution of PC1, PC2, and PC3 scores for selected Michigan counties (Wayne, Alger, Jackson, Sanilac, Isabella). Differences highlight: (1) Urban vs. lake/forest moderation on PC1 thermal amplitude; (2) Phase and peak shifts in PC2 (moisture accumulation) due to land--lake contrast and evapotranspiration; (3) Sharp PC3 impulses (turbulent mixing + radiative heating) preceding or accompanying potential convective or gust-driven outage windows.