Dual-Horizon Parking Availability Prediction with Multi-Factor Recommendation: An Empirical Study on Real-World Data

Abstract

To address the two key problems in urban parking guidance—missing real-time information and inaccurate short-term prediction—this paper presents an empirical study of dual-horizon availability prediction and multi-factor recommendation on the public Birmingham car park dataset. For 30-minute and 60-minute prediction horizons, we construct a 39-dimensional feature set comprising 38 time-series features in three groups—temporal, lag, and sliding-window statistics—plus one static capacity feature, and systematically compare LightGBM, XGBoost, LSTM, and GRU models. Results show that, with strong feature engineering, gradient-boosted trees significantly outperform recurrent neural networks (MAPE 5.53% versus 9.38% at the 30-minute horizon) and degrade the slowest as the horizon lengthens. Ablation experiments further reveal a horizon effect: short-horizon prediction relies most on sliding-window statistics, whereas at the longer horizon periodic temporal encodings become the dominant signal while lag features contribute almost nothing. Building on these predictions, we propose a prediction-driven multi-factor scoring recommendation strategy and evaluate it in a stylized simulation with data latency. The strategy achieves one-shot success rates of 98.8%, 98.4%, and 97.3% under normal, high-load, and “high-load + latency + strict threshold” scenarios respectively—on par with ideal real-time recommendation, significantly better than distance-only ranking, and without requiring a real-time data stream. Finally, we present a complete engineering implementation covering prediction models, recommendation decisions, and a user feedback loop.