Operational question
For grid operators, the value is not only knowing the future load trajectory, but also assessing how that trajectory would change under different news-event conditions.
News-aware counterfactual load analysis
CounterfactualLoad Forecasting
News-reported social, environmental, and grid events can reshape electricity demand beyond regular historical, weather, and calendar patterns. NACF treats news as an intervenable semantic treatment and estimates how the forecasted load trajectory changes when observed events are removed or anticipated scenarios are injected under the same historical context.
355 MWDay-ahead MAE on the 48 to 48 setting
511 MWDay-ahead RMSE in the original demand scale
13.6%-21.3%MAE improvement over the strongest baselines
-233 MWMean observed-news minus no-news-baseline comparison
Motivation
Background and Motivation
Modern power systems are increasingly affected by non-periodic social and operational shocks, including public health emergencies, emergency restrictions, policy interventions, and infrastructure disruptions. Many of these events are first reported, updated, and interpreted through news, making news text important for understanding event-driven demand changes.
Limit of factual prediction
Existing news-augmented load forecasting studies use news as an auxiliary feature to reduce point-forecast error, but they cannot answer what the load trajectory would be if the news condition were different.
Counterfactual load forecasting
Counterfactual analysis estimates the demand perturbation associated with removing an observed event or injecting an anticipated event scenario while keeping the same historical operating context.
Observed-confounding adjustment
The same high temperatures that drive cooling demand can also trigger weather-warning news. NACF therefore learns a confounder representation and balances it across treatment-intensity groups.
Motivation figure: factual forecasting optimizes the observed load trajectory, whereas NACF changes the news treatment while holding the historical context fixed.
Theoretical Foundations
Counterfactual Error Is Unobserved
Counterfactual outcomes are never observed, so the counterfactual prediction error cannot be directly minimized from data. NACF formalizes news as a continuous treatment in a potential-outcome framework and uses a continuous-treatment generalization bound to connect this unobservable error with observable training objectives.
Observational sample
Each sample contains the historical multivariate time series, the continuous semantic treatment encoding news information, and the future load trajectory.
X includes load, meteorological variables, and calendar indicators; T encodes news event information; Y is the future load horizon.
Counterfactual comparison
The target comparison changes the news treatment while holding the same historical operating context fixed.
This estimates news-related demand perturbations under observed versus alternative news conditions, such as the no-news baseline or an injected event scenario.
Why balance is needed
A model trained only on factual outcomes may exploit confounded correlations. For example, high temperatures can both trigger weather-warning news and directly increase cooling demand.
NACF learns a confounder representation Z and reduces its distributional discrepancy across treatment-intensity groups.
Generalization bound for continuous treatments
The bound decomposes the unobservable counterfactual error into terms that can be controlled with observed data.
Weighted factual loss maps to learned sample reweighting; IPM_Δ(Φ) maps to MMD-based representation balance; the discretization term maps to treatment-intensity binning for continuous news treatments.
Paper
Contributions
The paper proposes the News-Aware Counterfactual Load Analysis Framework (NACF), treating news not merely as an auxiliary predictive signal but as a continuous semantic treatment that can be intervened on for counterfactual load analysis.
1Counterfactual formulation for news-impact load analysis
2Structured news-to-treatment pipeline using offline LLM extraction
3Generalization-bound-guided architecture and training objective
4Empirical validation for reliable counterfactual load analysis
NACF Method
News-Aware Counterfactual Load Analysis Framework
Guided by the continuous-treatment bound, NACF converts raw news into structured event streams, learns confounder and treatment representations, and estimates treatment-dependent load trajectories with a varying-coefficient response network trained by weighted factual loss and IPM-based representation balance.
LLM-structured event stream
An offline LLM-based extraction step converts unstructured news articles into time-stamped event records with category, temporal type, scope, relevance, and textual evidence.
Confounder and treatment encoders
Historical load, meteorological variables, and calendar indicators are encoded as the adjustment representation, while the event stream is mapped into a continuous semantic treatment space.
Weighted factual loss and IPM balance
Learned sample reweighting controls the weighted factual prediction term, while MMD-based balance regularization reduces representation imbalance across treatment-intensity groups.
Counterfactual load analysis
The response network generates factual and counterfactual load trajectories under observed, removed, or injected news treatments to estimate demand perturbations.
Method overview: structured event extraction, representation learning, and counterfactual inference for load analysis.
Results
Forecasting and Diagnostics
The plots below summarize the main diagnostics reported in the published paper and the data behind the interactive examples. NACF maintains strong factual forecasting performance while showing treatment-intensity structure, improved representation balance, and interpretable news-related demand perturbations.
Category-specific perturbations
News effects are aggregated by category using absolute CATE and ITE distributions, showing heterogeneous impacts across environmental, social, economic, and grid events.
Ablation study
Removing reweighting or IPM regularization increases RMSE, especially in the day-ahead and two-day settings.
Balance-accuracy trade-off
The selected IPM weight sits near the practical elbow between prediction loss and representation balance.
Treatment intensity
The learned treatment representation preserves a graded notion of semantic departure from the no-news baseline.
Representation balance
The full model substantially reduces treatment-group discrepancy in the learned operating-context representation.
Synthetic probes
Controlled event texts produce differentiated response surfaces while holding the historical operating context fixed.
Interactive Demo
Counterfactual Scenario Explorer
This panel follows the tutorial notebook: observed-news prediction, no-news-baseline counterfactual prediction, and custom-news scenario prediction for the same historical operating context.
Tutorial window: Sydney sporting events
Forecast starts 2019-10-20 23:30:00. The comparison holds the same load, weather, and calendar history while changing the news treatment representation.
Forecast under observed news (factual)Forecast under no-news baseline (counterfactual)True load (ground truth)
Observed news in this window
Inaugural Rugby League World Cup Nines at Bankwest Stadium (Sydney) drew crowds of ~12,528 (Fri) and ~15,684 (Sat), creating concentrated event-driven electricity demand at the stadium and nearby precinct.
The Everest horse race at Royal Randwick (Randwick, NSW) attracted a reported crowd of 40,912 on Everest Day, representing a major spring racing event that increases local electricity demand for lighting, HVAC, catering and broadcast operations.
Notebook scenario summary
Exported from code/notebooks/counterfactual_inference_tutorial.ipynb: factual prediction with observed news, no-news-baseline counterfactual prediction, and custom-news scenario prediction under one shared historical context.
Case Study
COVID-19 Lockdown Perturbation
The published paper case study analyzes March 19 to 22, 2020 in NSW. Lockdown-related restrictions, isolation behavior, and economic shutdown signals dominate a concurrent heatwave context and yield a sustained negative model-estimated perturbation.
Case-study figure: factual forecast, counterfactual forecast with news masked, true future load, net perturbation, and high-relevance news signals for the same historical window.