Learning to Cool: State Space Models for Smarter Automobile AC Systems

We investigate the application of modern state-space models for time series forecasting of automobile air conditioning (AC) power draw, a task critical for improving heating, ventilation, and air conditioning (HVAC) control systems impacting fuel or battery consumption by up to 20% (Vasile-Müller, 2011). Traditional long short-term memory (LSTM) based approaches struggle with long-range dependencies in high-resolution vehicle telemetry, motivating our exploration of the Structured State Space Sequence (S4) architecture. We train and evaluate S4 forecaster variants on more than 7,000 sliding windows of data, then compare them against a hybrid LSTM-Attention baseline. Across multiple window configurations, S4 consistently outperforms the baseline, achieving a mean squared error (MSE) as low as $0.000288$, an $R^2$ up to $0.9820$, and a weighted mean absolute percentage error (WMAPE) as low as $0.0247$. We also report latency and parameter counts, showing that the S4 forecaster remains compact and suitable for in-vehicle deployment. These results demonstrate that structured state-space models can improve predictive accuracy for automotive HVAC forecasting while maintaining a lightweight footprint, providing a foundation for next-generation energy-aware control systems.