Dynamic Urban Air-Quality Forecasting through Multi-Modal Machine Learning on Transport, Weather, and Sensor Data

Abstract

Urban air pollution is a global challenge that urgently requires effective and timely solutions. City governments must allocate resources to protect their citizens from the adverse effects of air pollution while being optimally effective and cost-efficient. A growing number of low-cost air-quality sensors may help with timely intervention. However, they typically deliver low accuracy and sparse coverage. A novel forecasting approach is presented that leverages existing transport and weather data to develop an easy-to-implement model capable of producing dense air-quality estimates at multiple horizons. In addition, contradictory space–time behavior is investigated and a multi-modal forecasting model is proposed. By capturing short-term predictions and longer-term trends, the model effectively incorporates spatio-temporal correlations across different time scales. The approach is validated on sensor data from London, indicating that a multi-modal model combining transport, weather, and sensor data outperforms pure multi-task or single-task setups. Well-calibrated, short-term, and contradicting forecasts enable practical air-quality forecasting when an abundance of noisy sensor measurements is available. Cities can leverage existing data and create easy-to-implement, low-cost, dynamic air-quality forecasting solutions, facilitating timely interventions and resource allocation.

Forecasting air-quality measures at arbitrary spatio-temporal scales is a challenging task. Although Low-Cost Sensors (LCSs) have recently offered the opportunity to improve prediction quality by supplying inexpensive and often expressed spatio-temporal forecasts, they come with significant downsides: relatively low accuracy and limited coverage. A general multi-modal forecasting framework capable of providing LCS-based predictions at multiple time horizons is thus highly attractive, as it allows the combination of high-accuracy and low-cost measurements while addressing their individual downsides.