結合搜尋引擎關鍵字趨勢與多時間特徵預測特殊活動期間車站級客流量

Integrating Search Engine Index Trends and Multi-temporal Features for Predicting Bus Station-level Passenger Flow during Special Events

本研究以臺中市七期某百貨周年慶期間為例，開發堆疊模型預測公車站點的客流量，並蒐集該活動期間的搜尋引擎趨勢、票證資訊及其他相關的時間與外部因素數據。研究共考量19項特徵，並進行重要性表現分析。結果顯示，對客流量預測最重要的五項特徵為「前30分鐘客流量」、「前2週同時段平均客流量」、「前1天同時段客流量」、「前3週同時段平均客流量」及「前4週同時段平均客流量」。模型最終預測結果顯示，當使用所有特徵進行預測時，模型效能最佳，平均絕對誤差(Mean Absolute Error, MAE)為14.64，R2達0.86，處理時間為319秒。當僅使用前五大重要特徵時，MAE增加至17.51。若在前五大特徵基礎上依序加入其他特徵，其中「搜尋引擎關鍵字搜尋量」的加入對模型效能改善最大，MAE降低至15.17，且處理時間顯著減少至143秒。研究結果顯示，使用所有特徵雖可達最佳預測精度，但處理時間相對較長；若計算資源有限，則在前五大重要特徵基礎上加入「搜尋引擎關鍵字搜尋量」能在預測精度與計算效率間取得良好平衡，為特殊活動期間的有效替代方案。

This study developed a stacking model to predict passenger flow at a bus station in Taichung’s 7th Redevelopment Zone during the anniversary sale of a department store in the area. The study collected data on search engine index trends, ticketing information, and other relevant temporal and external factors corresponding to the sale period. A total of 19 features were considered, and an importance-performance analysis was conducted. The analysis results revealed that the top five most important features for passenger flow prediction were“previous_30_minutes_volume,”“same_time_2_weeks_ago_average_volume,”“same_time_previous_day_volume,”“same_time_3_weeks_ago_average_volume,”and“same_time_4_weeks_ago_average_volume.”The final prediction results indicated that when all considered features were selected, the developed model exhibited optimal performance, with its mean absolute error (MAE) and R2 value being 14.64 and 0.86, respectively, and a processing time of 319 seconds. When only the five most important features were used in the model, the MAE increased to 17.51. Therefore, other features were added in sequence to the top five features. The addition of“search engine index volume”caused the greatest improvement in the model performance, with the MAE reducing to 15.17 and the processing time significantly reduced to 143 seconds. The results of this study indicate that while using all features provides the best prediction accuracy, it also requires the longest processing time. If computational resources are limited, adding the“search engine index volume”to the top five most important features provides a good balance between prediction accuracy and computational efficiency during special events, making it a viable alternative.