Predicting Public Health Impacts of Electricity Usage

AI summary

Overview

• Research area: AI/ML applied to energy systems and public health — specifically, a domain-specific model that connects electricity consumption to monetized public health outcomes.
• Technical level: Intermediate. The paper is readable for those with some ML background (Transformers, LSTMs, MLPs, loss functions), but the health-economics side is explained from first principles.
• Scope: The paper introduces HealthPredictor, an end-to-end pipeline combining a fuel mix predictor, an air quality converter, and a health impact assessor, validated on three U.S. power regions and a case study on electric vehicle charging.

What This Paper Is About

Fossil-fuel power plants emit air pollutants that harm public health, but the chain from "electricity consumed" to "health damage" is complex, spanning fuel choice, emissions, atmospheric dispersion, and epidemiology. The authors build a single AI pipeline that predicts public health damages in dollars per megawatt-hour ($/MWh) directly from historical fuel mix data, so that the health signal can be embedded in demand-side decisions. The goal is to give users and system operators an actionable, health-based signal they can respond to, for example by shifting EV charging to cleaner hours.

Key Contributions

1. An end-to-end domain-specific AI model (HealthPredictor) that links electricity usage to public health outcomes through three modules: a fuel mix predictor, an air quality converter, and a health impact assessor. The paper states that prior work typically addressed only isolated pieces of this chain.
2. A health-informed loss function (Eq. 3 and Eq. 4) that trains the fuel mix predictor jointly on fuel mix accuracy and predicted health impact, weighted by a hyperparameter beta. This yields two operating regimes the authors call Fuel-mix-driven Opt (beta near 1, up to 0.998) and Health-driven Opt (smaller beta).
3. A released dataset and codebase linking hourly fuel mix compositions to internal and external health costs across all 67 U.S. balancing authorities for the latest available year, totaling 586,920 data points, available at https://github.com/Ren-Research/Health-Impact-Predictor.
4. A case study on health-aware EV charging that uses the predicted $/MWh signal to schedule charging and compares it against alternative scheduling strategies.

Main Findings

• Health-driven optimization lowers health prediction error: Sweeping beta across (0, 1) across CISO, ERCO, and PJM, the authors report that Health-driven Opt consistently achieves lower health impact NMAE than Fuel-mix-driven Opt. Figure 3 is cited for the trade-off; specific numerical NMAE values are not reported in the provided text.
• Transformers beat LSTMs: Transformer-based architectures consistently outperform LSTM baselines across both optimization objectives and both prediction windows (T = 24 and T = 72 hours).
• Downstream health signals must be trained in: The authors state that incorporating the downstream health impact into the predictor is necessary to produce accurate signals for health-informed energy management. They note beta cannot be set to 1, because doing so would prevent the model from learning air dispersion or health outcomes.
• EV charging can be shifted for health benefit: In simulation across CISO, PJM, and ERCO, the optimal schedule described in Section 6 reduces total health impacts by roughly 24–42% versus the First Hours and Latest Hours strategies, and roughly 15–20% versus Continuous Charging. The conclusion instead states potential reductions of 17–42% compared to other charging strategies.
• Health damages are material in scale: Coal-fired power plants' PM2.5 emissions are cited as having caused approximately 460,000 excess deaths between 1999 and 2020. Outdoor air pollution is described as the second largest risk factor for noncommunicable diseases, contributing to approximately 4.2 million premature deaths globally each year, and a 2024 European Environment Agency assessment cited in the paper puts power-generation air pollution damages at roughly 1 percent of GDP in Europe.
• Fossil fuels persist in the grid: The U.S. EIA projects that even by 2050 coal power generation remains around 180 billion kWh under an alternative scenario where plants operate under rules existing before early 2024.

Methodology in Plain English

The pipeline has three stages, trained together rather than in isolation.

1. Fuel mix predictor. The model takes a sequence of historical hourly fuel mix data (the share of generation from sources such as coal, oil, and gas) and predicts the mix for the next time horizon. The authors chose a Transformer architecture because it captures long-range dependencies in time-series data better than alternatives; they also built LSTM variants as baselines.
2. Air quality converter. The predicted fuel mix is multiplied by per-fuel emission factors to estimate pollutant emissions. For dispersion, the authors follow the simplified approach used in the EPA's COBRA tool, using the prevailing weather pattern, and approximate the dispersion transformation with a custom neural network layer. That layer takes source pollutant quantities and environmental features as input and predicts pollutant quantities at receptor regions.
3. Health impact assessor. Changes in pollutant concentration at each receptor are converted into adverse health outcomes using a log-linear concentration-response model (with linear models used in some cases), then monetized using frameworks such as the value of a statistical life. The output is expressed as dollars per megawatt-hour. The paper notes that health impacts from exposure at time t unfold over subsequent years, typically within a five-year window.

The whole pipeline is trained with a loss that is a weighted sum of fuel mix prediction error and health impact prediction error, with beta controlling the balance. In experiments, that error is split evenly (1−beta)/2 between within-region ("internal") health impact and outside-region ("external") health impact.

Data: Hourly fuel mix data from the U.S. EIA; health impact values based on the EPA's AVerted Emissions and geneRation Tool (AVERT). The experiments cover three regions: California (CISO), Texas (ERCO), and the Mid-Atlantic (PJM). There are six input features (fuel mix percentages and time period) and two output features (internal and external health impacts). For the EV case study, the authors use ACN-Data for charging details and the National Household Travel Survey for arrival and departure time distributions.

Why This Matters

Impact on research: The paper argues that prior work either assessed health impacts from pollutants or modeled energy-to-emissions conversion, but rarely connected demand-side electricity use directly to health outcomes. HealthPredictor is presented as an end-to-end alternative, and the released dataset of 586,920 data points across 67 U.S. balancing authorities addresses what the authors describe as fragmented and dispersed data. It also connects to the emerging field of health-informed computing.

Real-world applications:

• EV charging scheduling: Shifting charging to hours with lower predicted health damages from generation, which the case study quantifies across three U.S. regions.
• Data center workload scheduling: The paper frames AI and large language model growth as a driver of unprecedented data center energy demand, and positions health signals as a signal for flexible load management.
• Energy efficiency programs and renewable siting: The authors note EPA annual-average health damages are already used to inform such decisions, a use case the finer temporal resolution of this model could extend.
• Individual consumer guidance: Providing users a real-time $/MWh signal showing the health cost of using electricity at a given time.

Industry relevance: The paper positions the model as filling gaps left by existing tools. The EPA's reported health damages lack temporal variation, which limits their use for dynamic demand-side management. WattTime provides real-time health signals, but the paper states these reflect only marginal damages and that the underlying methodology is proprietary, limiting external verification. HealthPredictor is offered as a transparent, region- and fuel-diverse alternative.

Future Directions

• Longer prediction horizons: The authors list short time windows as a limitation, noting their predictions do not extend to long-term scenarios.
• Reducing dispersion uncertainty: The paper acknowledges potentially high uncertainty in air dispersion due to complex interactions between emission sources and meteorological conditions, even though the EPA's dispersion model is used as ground truth.
• Richer dispersion modeling: The authors state that more advanced models incorporating real-time weather conditions are possible beyond their simplified prevailing-weather approach.
• Broader generalizability: The empirical results cover three balancing authorities; the released dataset spans all 67 U.S. balancing authorities, raising the open question of how well the approach transfers across the full set of regions and fuel mixes.

Target Audience

Researchers and practitioners working at the intersection of machine learning, energy systems, and environmental health — particularly those interested in demand-side management, sustainability-aware computing, and health-informed scheduling of flexible loads such as EV charging and data center workloads. It is also useful for policymakers and energy analysts who need to understand the trade-offs between grid-level health impacts and system-level fuel mix predictability, and for ML researchers looking for a concrete example of embedding a societal outcome directly into a training objective.

Authors’ abstract

The electric power sector is a leading source of air pollutant emissions, impacting the public health of nearly every community. Although regulatory measures have reduced air pollutants, fossil fuels remain a significant component of the energy supply, highlighting the need for more advanced demand-side approaches to reduce the public health impacts. To enable health-informed demand-side management, we introduce HealthPredictor, a domain-specific AI model that provides an end-to-end pipeline linking electricity use to public health outcomes. The model comprises three components: a fuel mix predictor that estimates the contribution of different generation sources, an air quality converter that models pollutant emissions and atmospheric dispersion, and a health impact assessor that translates resulting pollutant changes into monetized health damages. Across multiple regions in the United States, our health-driven optimization framework yields substantially lower prediction errors in terms of public health impacts than fuel mix-driven baselines. A case study on electric vehicle charging schedules illustrates the public health gains enabled by our method and the actionable guidance it can offer for health-informed energy management. Overall, this work shows how AI models can be explicitly designed to enable health-informed energy management for advancing public health and broader societal well-being. Our datasets and code are released at: https://github.com/Ren-Research/Health-Impact-Predictor.