基于Mamba改进的停车可用率时序预测研究

Research on Time Series Prediction of Parking Availability Based on Improvements to Mamba

摘要: 实现精准的停车预测，是提高停车场使用效率、实现停车资源合理配置的重要内容。现有停车可用率预测方法在停车时序特征的挖掘方面有待进一步提升，因此本文分析了区域停车可用率的时空特征，提出了Mamba-ATS (Mamba-Attention Time Series Forecasting)模型。模型首先融合了停车可用率的时间序列数据和停车场的属性数据；然后使用Mamba模块提取长期变化特征，再引入注意力机制捕获短期变化特征，最后通过自适应权重调整模块融合长短期特征并输出预测结果。使用上海市区域停车场数据验证了Mamba-ATS模型的有效性，相较于单独使用Mamba模型以及其它基于注意力机制的模型，提出模型的平均绝对误差MAE降低了13%，均方误差MSE降低了26%，总体上具有较好的预测效果。

Abstract: Accurate parking prediction is crucial for enhancing the efficiency of parking lot usage and achieving optimal allocation of parking resources. Current methods for predicting parking availability have room for improvement, especially in mining the characteristics of parking time-series data. Therefore, this paper analyzes the spatiotemporal features of regional parking availability and proposes the Mamba-ATS (Mamba-Attention Time Series Forecasting) model. The Mamba-ATS model first integrates time-series data on parking availability with attribute data from parking lots. It then employs the Mamba module to extract long-term trend features, while incorporating an attention mechanism to capture short-term fluctuation features. Finally, it fuses these long-term and short-term features through an adaptive weight adjustment module, thereby generating the prediction output. The effectiveness of the Mamba-ATS model was validated using data from parking lots in Shanghai. Compared to using the Mamba model alone and other models based on the attention mechanism, the proposed model demonstrates a 13% reduction in Mean Absolute Error (MAE) and a 26% reduction in Mean Squared Error (MSE). Overall, it achieves superior predictive performance.