To address the problem of frequent fluctuations and insufficient prediction stability in the number of available parking spaces, a short-term forecasting method based on Crested Porcupine Optimization (CPO), Variational Mode Decomposition (VMD), and a Dual-Weighted Long Short-Term Memory network (DW-LSTM) is proposed. First, the key parameters of VMD are adaptively optimized using the CPO algorithm to achieve a stable multi-scale decomposition of the parking availability time series. Then, adaptive component weighting and temporal weighting mechanisms are proposed to dynamically adjust the contributions of decomposed components and significant historical time steps. Finally, the LSTM network is employed to perform multi-step forecasting of parking availability. Experimental results based on real-world parking availability data from Hangzhou show that the proposed method outperforms comparative methods in terms of mean absolute error, root mean square error, and mean absolute percentage error. Moreover, the method maintains stable prediction performance in cross-dataset forecasting tasks, which verifies its effectiveness in complex parking space time series prediction.