Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Causal Feature Selection for Weather-Driven Residential Load Forecasting

Elise Zhang, François Mirallès, Stéphane Dellacherie, Di Wu, Benoit Boulet

TL;DR

Weather strongly drives residential electricity demand, but naïvely adding many meteorological covariates can inflate models and harm generalization. The paper evaluates PCMCI-based causal feature selection against a non-causal MI filter and no-selection baselines across GRU, TCN, and PatchTST models in a Southern Ontario city-scale STLF case study, using IESO load and ERA5 weather data. It finds that selecting direct lagged causal drivers (F3) yields a compact feature set and often improves accuracy while enhancing robustness under extreme weather, compared to larger feature sets. The approach is model-agnostic and practically useful for day-ahead planning, as it reduces input redundancy and improves resilience to distribution shifts.

Abstract

Weather is a dominant external driver of residential electricity demand, but adding many meteorological covariates can inflate model complexity and may even impair accuracy. Selecting appropriate exogenous features is non-trivial and calls for a principled selection framework, given the direct operational implications for day-to-day planning and reliability. This work investigates whether causal feature selection can retain the most informative weather drivers while improving parsimony and robustness for short-term load forecasting. We present a case study on Southern Ontario with two open-source datasets: (i) IESO hourly electricity consumption by Forward Sortation Areas; (ii) ERA5 weather reanalysis data. We compare different feature selection regimes (no feature selection, non-causal selection, PCMCI-causal selection) on city-level forecasting with three different time series forecasting models: GRU, TCN, PatchTST. In the feature analysis, non-causal selection prioritizes radiation and moisture variables that show correlational dependence, whereas PCMCI-causal selection emphasizes more direct thermal drivers and prunes the indirect covariates. We detail the evaluation pipeline and report diagnostics on prediction accuracy and extreme-weather robustness, positioning causal feature selection as a practical complement to modern forecasters when integrating weather into residential load forecasting.

Causal Feature Selection for Weather-Driven Residential Load Forecasting

TL;DR

Weather strongly drives residential electricity demand, but naïvely adding many meteorological covariates can inflate models and harm generalization. The paper evaluates PCMCI-based causal feature selection against a non-causal MI filter and no-selection baselines across GRU, TCN, and PatchTST models in a Southern Ontario city-scale STLF case study, using IESO load and ERA5 weather data. It finds that selecting direct lagged causal drivers (F3) yields a compact feature set and often improves accuracy while enhancing robustness under extreme weather, compared to larger feature sets. The approach is model-agnostic and practically useful for day-ahead planning, as it reduces input redundancy and improves resilience to distribution shifts.

Abstract

Weather is a dominant external driver of residential electricity demand, but adding many meteorological covariates can inflate model complexity and may even impair accuracy. Selecting appropriate exogenous features is non-trivial and calls for a principled selection framework, given the direct operational implications for day-to-day planning and reliability. This work investigates whether causal feature selection can retain the most informative weather drivers while improving parsimony and robustness for short-term load forecasting. We present a case study on Southern Ontario with two open-source datasets: (i) IESO hourly electricity consumption by Forward Sortation Areas; (ii) ERA5 weather reanalysis data. We compare different feature selection regimes (no feature selection, non-causal selection, PCMCI-causal selection) on city-level forecasting with three different time series forecasting models: GRU, TCN, PatchTST. In the feature analysis, non-causal selection prioritizes radiation and moisture variables that show correlational dependence, whereas PCMCI-causal selection emphasizes more direct thermal drivers and prunes the indirect covariates. We detail the evaluation pipeline and report diagnostics on prediction accuracy and extreme-weather robustness, positioning causal feature selection as a practical complement to modern forecasters when integrating weather into residential load forecasting.

Paper Structure

This paper contains 22 sections, 1 equation, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Feature Selection for Load Forecasting
  4. Causality and Causal Feature Selection
  5. Problem Formulation
  6. City-Level Short-Term Load Forecasting
  7. Feature Selection
  8. Methodology
  9. PCMCI Causal Discovery for Feature Selection
  10. Non-Causal Feature Selection Baseline
  11. Evaluation Pipeline
  12. Data Split
  13. FS and STLF Setup
  14. Diagnostics
  15. Experiments
  16. ...and 7 more sections

Figures (3)

  • Figure 1: Markov Blanket: Parents, children, spouses suffice for prediction under Causal Markov + Faithfulness
  • Figure 2: Administrative regions of Southern Ontario from the census divisions of the https://www12.statcan.gc.ca/census-recensement/2021/geo/sip-pis/boundary-limites/index2021-eng.cfm?year=21 by Statistics Canada. Map created with https://felt.com/map GIS platform.
  • Figure 3: Demo: Toronto's OOD Weather Events Identified in Test Year, Including Summer Heat and Winter Cold Wave, Heavy Precipitation