An Encoder-Enhanced Autoformer for Multi-Scale Parking Availability Forecasting

Parking space availability forecasting is challenging due to noise contamination, non-stationary dynamics, and complex multi-scale temporal dependencies. This study proposes a wavelet-enhanced Autoformer with dynamic multi-kernel decomposition for parking space availability forecasting. Discrete wavelet transform is used to enhance the input sequence by suppressing high-frequency disturbances while preserving the dominant temporal structure. Within the Autoformer framework, a dynamic multi-kernel decomposition module is developed by introducing multiple moving-average kernels with different scales and an input-dependent dynamic weighting mechanism, enabling the adaptive decomposition and fusion of trend and fluctuation components and thereby improving the representation of complex multi-scale temporal patterns. The proposed framework is evaluated using real parking operation data collected from the Liangshan Jingyuan parking lot in Chongqing, China, including 1248 observations with a 2 h sampling interval. Multi-step forecasting experiments are conducted under short-, medium-, and long-term prediction settings with different label-length configurations. Across different forecasting horizons and label-length settings, the proposed model consistently outperforms the original Autoformer, PatchTST, and a range of conventional and recent forecasting baselines. In particular, under the 24-step forecasting setting, the proposed method reduces MAE, RMSE, and WMAPE by approximately 10% relative to the original Autoformer. Moreover, the results suggest that the dynamic multi-kernel mechanism improves the model’s ability to adaptively capture scale-dependent temporal patterns across different forecasting tasks.