Forecasting heat-related impacts with multivariate multi-step time series models using advanced deep learning

• Advanced DL models are used to forecast heat-related morbidity across counties in Sweden. • N-BEATS for heat impact forecasting outperforms classic statistical and other DL models. • Individually trained N-BEATS highlights the localized nature of heatwave impacts. • Incorporating exogenous variables in N-BEATS improves prediction performance. Record-breaking temperatures and frequent heatwaves have been experienced worldwide in recent years. Heatwaves pose an escalating threat to public health and heat-related impact forecasting is critical to implementing suitable mitigation strategies. Deep learning (DL) models, notably Long Short-Term Memory (LSTM), have been widely applied for heat-related impact forecasting. However, the emergence of state-of-the-art forecasting DL architectures such as Neural Basis Expansion Analysis for Interpretable Time Series Forecasting (N-BEATS) provides a novel solution for long-term heat-related impact forecasting. This study develops, evaluates, and compares multiple time series forecasting models—including advanced DL architectures (N-BEATS, N-HiTS, LSTM), a classical statistical model (ARIMA), and a Naïve seasonal baseline—to predict heat-related morbidity across 21 Swedish counties using data from 2008 to 2023. Both local (individually trained) and global (cross-learning across counties) modeling strategies were explored, incorporating exogenous variables (Heatwave Index and number of people with respiratory disease), and comparing recursive and Multi-Input-Multi-Output (MIMO) forecasting output strategies. Results indicate that the local N-BEATS model achieves superior predictive performance, particularly when both exogenous variables are included. MIMO generally yields a better performance by mitigating error propagation over extended forecasting horizons. Moreover, individually trained N-BEATS models outperform cross-learning global N-BEATS, underscoring the importance of localized adaptation plans. These findings highlight the potential utility of multivariate N-BEATS for more accurate heatwave impact forecasting. This study can complement and support early warning frameworks by integrating the developed impact forecast model with existing hazard models, thereby enabling more proactive public health interventions and improving community resilience to heatwaves.