Predicting Public Health Impacts of Electricity Usage¶

Conference: NeurIPS 2025 arXiv: 2511.22031 Code: https://github.com/Ren-Research/Health-Impact-Predictor Area: Earth Science / AI for Social Good Keywords: Public Health, Power Systems, Air Pollution, End-to-End Prediction, EV Charging Optimization

TL;DR¶

This paper proposes HealthPredictor, an AI pipeline that maps electricity consumption end-to-end to public health damages (measured in $/MWh), comprising three modules: fuel mix prediction, air quality conversion, and health impact assessment. Health-driven optimization significantly reduces health impact prediction error compared to fuel-mix-driven baselines, and achieves a 24–42% reduction in health damages in an EV charging scheduling case study.

Background & Motivation¶

Background: The power sector is one of the primary sources of air pollutant emissions. PM2.5, SO2, NOx, and other pollutants generated by fossil fuel power generation affect public health through atmospheric transport. Even by 2050, fossil fuels are projected to retain a substantial share of U.S. electricity supply.

Limitations of Prior Work: - Existing studies either focus solely on epidemiological associations between air pollution and health, or on energy-to-emission conversion, lacking an end-to-end framework. - Annual average health damage estimates provided by the EPA lack temporal variability, making them unsuitable for dynamic demand-side management. - WattTime provides real-time health impact signals, but the methodology is opaque and reflects only marginal damages.

Key Challenge: A large fraction of electricity demand is dynamically controllable (e.g., data center loads, EV charging), yet AI tools that directly link electricity consumption decisions to specific health outcomes are absent.

Key Insight: Design an end-to-end pipeline that begins from fuel mix time series, proceeds through emission estimation and atmospheric dispersion modeling, and ultimately outputs a health cost signal in $/MWh via health economics valuation.

Core Idea: Directly optimize the fuel mix predictor using health impact loss (rather than optimizing only fuel prediction accuracy), so that the prediction signal more faithfully reflects real-world health consequences.

Method¶

Overall Architecture¶

A three-stage pipeline: Fuel Mix Predictor (Transformer-based time series model) → Air Quality Converter (emission estimation + neural network dispersion model) → Health Impacter (log-linear concentration-response model + economic valuation). During end-to-end training, health impact loss is backpropagated to the fuel mix predictor.

Key Designs¶

Fuel Mix Predictor:
- Function: Predicts the fuel usage proportions for the next $T$ hours (24h/72h) based on historical fuel mix time series.
- Mechanism: Transformer architecture (single-layer encoder-decoder, 4-head attention), with inputs embedded into a 64-dimensional space.
- Design Motivation: Fuel mix data exhibits long-range temporal dependencies (diurnal cycles, seasonality), which Transformers are well-suited to capture.
Air Quality Converter:
- Function: Computes emissions from fuel mix predictions and models the atmospheric dispersion of pollutants.
- Mechanism: Emission estimation is based on per-fuel emission factors; dispersion modeling uses a 3-layer MLP to approximate the source-receptor matrix of EPA's COBRA model, i.e., $\mathscr{P}_r^1, ..., \mathscr{P}_r^M = D_{\boldsymbol{w}}(\mathscr{P}_s)$.
- Design Motivation: Full physical dispersion models such as COBRA are computationally expensive (5–20 minutes per run); the neural network surrogate enables real-time prediction.
Health Impacter:
- Function: Converts changes in pollutant concentrations into monetized health damages ($/MWh).
- Mechanism: A log-linear concentration-response model $\Delta Y^i = Y_0^i \times \text{POP}^i \times (1 - e^{-\alpha^T \Delta \mathscr{P}_r^i})$, where $Y_0^i$ is the baseline incidence rate, $\text{POP}^i$ is the exposed population, and $\alpha$ is the epidemiological coefficient.
- Distinguishes between internal (within the BA region) and external (outside the BA region) health impacts.

Loss & Training¶

Health-driven loss: $\mathcal{L} = \beta \|y_t - \hat{y}_t\|^2 + (1-\beta) \|y_{\text{impact},t} - g(\hat{y}_t, I)\|^2$

where $\beta$ controls the trade-off between fuel prediction accuracy and health impact prediction accuracy. $\beta$ close to 1 corresponds to fuel-driven optimization; smaller $\beta$ corresponds to health-driven optimization. Key insight: directly optimizing the health impact loss reduces downstream health prediction error more effectively than optimizing fuel prediction alone.

Key Experimental Results¶

Main Results¶

Evaluated on three major U.S. power regions — CISO (California), ERCO (Texas), and PJM (Mid-Atlantic) — covering diverse energy mixes and population densities.

Method	Health Impact NMAE (T=24)	Health Impact NMAE (T=72)
LSTM Fuel-driven	Higher	Higher
LSTM Health-driven	Moderate	Moderate
Transformer Fuel-driven	Moderate	Moderate
Transformer Health-driven	Lowest	Lowest

EV Charging Case Study¶

Charging Strategy	Health Damage Reduction
First Hours (earliest charging)	baseline
Latest Hours (latest charging)	baseline
Continuous Charging	15–20% reduction
Optimal (Ours)	24–42% reduction

Key Findings¶

Health-driven optimization consistently outperforms fuel-mix-driven optimization in health impact prediction accuracy.
The Transformer architecture comprehensively outperforms LSTM baselines.
Incorporating health impact into predictor training is a necessary condition for delivering accurate signals to downstream users.
Dataset construction is itself a major engineering contribution: 586,920 data points covering 67 U.S. Balancing Authorities (BAs).

Highlights & Insights¶

End-to-end health-aware training paradigm: Downstream health impacts are directly incorporated into the training objective of the upstream predictor, rather than optimizing each stage independently. This "objective alignment" philosophy is broadly applicable to multi-stage decision-making problems.
Dataset engineering: Cross-institutional data fusion (EIA's 8 fuel categories vs. EPA eGRID's 40+ categories; BA-level regions vs. county-level health data) is an underappreciated yet critically important contribution.
Strong practical operability: The pipeline provides real-time $/MWh health cost signals that can be directly consumed by end users or scheduling systems.

Limitations & Future Work¶

Coverage is limited to short-term prediction horizons (24–72h); long-term cumulative health effects are not modeled.
The atmospheric dispersion model is a simplified version of EPA COBRA, and does not account for real-time meteorological variability.
Indirect health impacts of greenhouse gas emissions are not considered.
The neural network dispersion model has limited physical interpretability.

vs. EPA COBRA: COBRA is a standard tool but is slow to run and requires cumbersome data inputs; HealthPredictor achieves real-time inference.
vs. WattTime: WattTime provides similar signals but with an opaque methodology; the proposed method is fully transparent and end-to-end trainable.
The work serves as a demonstration for the AI for Social Good community, showing how the externalities (health damages) of technical systems (electricity consumption) can be quantified and fed back into decision-making.

Rating¶

Novelty: ⭐⭐⭐⭐ The problem formulation and solution framework for end-to-end electricity-to-health mapping are highly novel.
Experimental Thoroughness: ⭐⭐⭐⭐ Three regions + EV charging case study + ablation studies provide comprehensive validation.
Writing Quality: ⭐⭐⭐⭐ Problem motivation is clear; pipeline description is complete.
Value: ⭐⭐⭐⭐⭐ Directly supports public health decision-making with substantial societal impact.