Causal Feature Selection for Weather-Driven Residential Load Forecasting

Elise Zhang 1 , François Mirallès 2 , Stéphane Dellacherie 3 4 , Di Wu 1 , Benoit Boulet 1 · 2025

Read Causal Feature Selection for Weather-Driven Residential Load Forecasting by Elise Zhang 1 , François Mirallès 2 , Stéphane Dellacherie 3 4 , Di Wu 1 , Benoit Boulet 1 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-20508.txt.

Introduction

Zhang, Mirallès, Dellacherie, Wu, and Boulet ask which weather data truly drives home electricity use. Their 2025 paper tests causal feature selection for load forecasting. Weather strongly drives residential demand for heating and cooling. Throwing in every weather variable can still hurt accuracy. A principled way to choose inputs is needed.

The case study covers Southern Ontario cities and regions. Hourly consumption data came from the provincial system operator. ERA5 reanalysis supplied the matching weather fields. Three forecasting models were tested: GRU, TCN, and PatchTST. Plain correlation-based selection favored radiation and moisture variables. The causal PCMCI method picked direct thermal drivers instead. It pruned indirect variables that merely correlate with demand. Causal selection kept feature sets compact across nine cities. Accuracy matched or improved against using everything. Robustness held during extreme cold, heat, and heavy rain. The module is model-agnostic and lightweight for operators.

Catalog record

CREATORElise Zhang 1 , François Mirallès 2 , Stéphane Dellacherie 3 4 , Di Wu 1 , Benoit Boulet 1
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE26 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-20508

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Causal Feature Selection for Weather-Driven Residential Load Forecasting Elise Zhang 1 , François Mirallès 2 , Stéphane Dellacherie 3 4 , Di Wu 1 , Benoit Boulet 1 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. I Introduction Short-term load forecasting (STLF) estimates near-future electricity demand (hours to days ahead) so operators can schedule generation, ensure reliability, and manage markets with minimal cost and risk. As grid-scale storage remains limited and costly, supply must closely track demand in real time; accurate STLF is therefore central to reliable and economical grid operations. At city level, STLF typically augments autoregressive load history with calendar signals (hour, weekday, holidays) and exogenous weather (temperature, humidity, cloud cover, radiative fluxes, precipitation). Prior studies [ 1 , 2 ] find that appropriate weather integration improves forecasting performance. Yet, adding too many meteorological covariates inflates the feature space, which adds to model complexity and confounds the model with spurious correlations. This might negatively impact model performance, especially under seasonal regime shifts. This paper studies whether causal feature selection (causal FS) helps isolate the most relevant weather drivers for electric load while reducing input dimensionality. We present a focused case study on Southern Ontario , pairing administrative-level residential electricity consumption from Independent Electricity System Operator (IESO)’s Hourly Consumption by FSA reports with ERA5 [ 3 ] hourly meteorological reanalysis data. We compare no selection, a non-causal filter, and PCMCI-based causal selection on single-city forecasting with GRU [ 4 ] , TCN [ 5 ] , and PatchTST [ 6 ] . Our contributions are summarized as follows: ( i ) A weather-informed load forecasting case study and evaluation pipeline that compares feature-selection regimes across forecasts on multiple cities; ( ii ) A feature-level analysis showing that causal FS favors direct thermal drivers consistent with domain mechanisms, whereas non-causal filtering retains more indirect correlates; ( iii ) we report diagnostics on model accuracy and robustness under extreme weather, positioning causal FS as a model-agnostic module for weather integration. The remainder of the paper is structured as follows: Section II introduces related work; Section III formulates the problem; Section IV presents the causal and non-causal selection modules and evaluation design; Section V presents the experimental results; and Section VI summarizes key takeaways and briefly discusses future work. II Related Work II-A Feature Selection for Load Forecasting STLF methods commonly combine load history with calendar features, and exogenous meteorology, reflecting the well-documented sensitivity of electricity demand to weather across seasons. Existing studies [ 1 , 2 ] consistently report that weather is an important exogenous driver of electric load, and that appropriate weather integration improves accuracy and reliability. As the feature space grows with exogenous drivers, selecting appropriate input features becomes critical for both model generalization and efficiency. Classical feature selection (FS) seeks to identify the minimal feature set that still exhibits optimal prediction performance. [ 7 ] FS approaches are often grouped into: (i) filters (e.g., maximal relevance/minimal redundancy, mRMR, using mutual information) [ 8 ] ; (ii) wrappers , which search subsets with a predictive model in the loop [ 9 ] ; and (iii) embedded methods that shrink/select during training (e.g., LASSO/Elastic Net) [ 10 ] . While these approaches can reduce error and complexity, they primarily exploit correlations with the prediction target rather than underlying mechanisms (true causal relations), risking spurious selections and brittleness under distribution shift. Power systems studies also report that using all features is rarely optimal, reinforcing the need for principled selection [ 11 ] . II-B Causality and Causal Feature Selection Fig. 1: Markov Blanket: Parents, children, spouses suffice for prediction under Causal Markov + Faithfulness Causality captures the true data-generating mechanisms of systems, not mere predictability. Causal FS aims to select features by formal causal analysis [ 12 ] and seek variables that are causally sufficient to explain the target variable, rather than merely correlated with it. Under the Causal Markov and the Faithfulness Assumptions [ 13 , 14 ] , a causal Directed Acyclic Graph (DAG) entails that each node X is conditionally independent of all its non-descendants, given its direct parents. In other words, the minimal subset of features that renders X independent of all remaining variable are those with a direct link with X (parents, children, and “spouses”). This subset is referred to as the Markov blanket (MB) of X (Fig. 1 ), and will be sufficient to explain X . Any feature outside of MB is redundant for predicting X , given MB. Consequently, selecting MB provides a minimal and causally sufficient feature set, aligning with the goal of FS. Classical work applies this principle directly to FS: KS algorithm [ 15 ] , which uses MB-based screening, was established as a criterion for optimal feature selection as early as 1996; subsequent work on MB induction (e.g., IAMB [ 16 ] ) operationalizes the discovery of MBs around a target at scale. For multivariate time series , the relevant blanket narrows to direct causal parents, both lagged and instantaneous, available during inference. Time series causal discovery approaches (e.g., TiMINo [ 17 ] , PCMCI/PCMCI+ [ 18 , 19 ] ) target exactly the identification of such parents, and provide a causality-aware lens for FS tailored to time series-related tasks. III Problem Formulation We consider hourly residential electricity demand for a set of geographically proximate cities/regions \mathcal{C} . For each city/region c\in\mathcal{C} , we observe historical load \mathcal{\textbf{Y}}^{(c)}_{0:t}=[y^{(c)}_{0},y^{(c)}_{1},y^{(c)}_{2},...,y^{(c)}_{t}] . Calendar-datetime features \mathcal{\textbf{D}}_{0:t} are also extracted (e.g., hour-of-day, day-of-week, month, day-of-year, holidays), as well as corresponding meteorological predictors \mathcal{\textbf{W}}^{(c)}_{0:t} (e.g., temperature, cloud cover, radiation, precipitation, etc.). Here, weather is treated as a causal driver of electricity demand and is assumed available at prediction time via observations. III-A City-Level Short-Term Load Forecasting Let \mathcal{P}^{(c)}=[\mathbf{D},\,\mathbf{Y}^{(c)},\,\mathbf{W}^{(c)}] denote the full predictor set. Given a lookback window L and forecast horizon H , we pose forecasting as supervised learning from historical predictors to future demand. In this study we use one week lookback and one day horizon, i.e., L{=}168 hours and H{=}24 hours. For city c , define its predictor history \mathcal{P}^{(c)}_{\mathrm{hist}}=[\mathbf{D}_{(t-L+1):t},\,\mathbf{Y}^{(c)}_{(t-L+1):t},\,\mathbf{W}^{(c)}_{(t-L+1):t}] and target \mathbf{Y}^{(c)}_{\mathrm{tgt}}=\mathbf{Y}^{(c)}_{(t+1):(t+H)} . The task is, for each c , to learn a function f^{(c)}:\mathcal{P}^{(c)}_{\mathrm{hist}}\mapsto\mathbf{Y}^{(c)}_{\mathrm{tgt}} that predicts

Read the complete text