整合外叫人流資訊估計與派車決策之電梯派車最佳化問題

Abstract

電梯派車系統面臨的核心挑戰在於外部呼叫資訊的不確定性，傳統的電梯派車方法僅能依據乘客按下的上行或下行按鈕進行派車，無法準確掌握等候乘客的實際人數以及其目的地樓層。這種資訊不足的情況可能導致電梯到達時發現等候乘客人數超出容量限制，或是無法有效規劃行程路線，進而造成額外的電梯派遣需求與更長的等候時間。隨著建築物樓層數量的增加及乘客流量模式的多樣化，傳統派車演算法的效率有限，因此需要發展能夠整合乘客流量估計與即時派車決策的智慧化電梯派車系統。

本研究提出一個整合性框架，結合人流資訊估計與最佳化派車演算法來解決上述問題。在人流資訊估計的部分，我們實作並比較了多種方法來估計乘客的數量和目的地，包括簡單頻率法、MV10 演算法 (Utgoff and Connell, 2012) 和隨機建模 (Sorsa et al., 2018) 等等。在派車決策的部分，我們建構了一個多目標最佳化模型，以最小化大於閾值的乘客等待時間與電梯能耗為目標，並採用基因演算法求解不同交通情境下的電梯派車問題。

實驗結果證實，當乘客需求呈現高變異性或集中分布特性時，估計方法展現出最顯著的效益，特別是在乘客的起點非常集中的情境下，簡單頻率法能將估計誤差從 8.294 降低至 0.650。比較不同估計方法時，MV10 演算法在資料稀少的初期階段表現較為穩定，而簡單頻率法在累積充足歷史資料後具有較佳的準確性。使用真實資料的實驗進一步驗證了歷史資料初始化的重要性：使用 28 天的資料訓練估計模型，並根據資料量動態切換估計方法，目標式值優於傳統無估計方法，證明了整合即時人流資訊估計的派車系統在實際應用中的有效性與實用價值。

The core challenge faced by elevator dispatching systems lies in the uncertainty of outer call information. Traditional elevator dispatching methods can only assign elevators based on passengers pressing up or down buttons, without accurately knowing the actual number of waiting passengers or their destination floors. This lack of information may lead to situations where elevators arrive to find that the number of waiting passengers exceeds capacity limits, or fail to effectively coordinate elevator routing, resulting in additional dispatches and increased waiting times. With an increasing number of floors in buildings and greater variability in passenger flow patterns, the limitations of traditional dispatching algorithms become increasingly evident. Therefore, it is essential to develop intelligent elevator dispatching systems that integrate passenger flow estimation with real-time dispatch decisions.

This study proposes an integrated framework that combines traffic flow estimation with optimized dispatching algorithms to address the aforementioned challenges. In terms of flow estimation, we implement and compare multiple methods for estimating the numbers and destinations of passengers, including the naive frequency method, the MV10 algorithm (Utgoff and Connell, 2012), and stochastic modeling approaches (Sorsa et al., 2018). For dispatching decisions, we construct a multi-objective optimization model aiming to minimize passenger waiting time exceeding a threshold and elevator energy consumption. The genetic algorithm is employed to solve elevator dispatching problems under various traffic scenarios.

Experimental results confirm that when passenger demand exhibits high variability or concentrated distribution characteristics, estimation methods demonstrate the most significant benefits, particularly in scenarios where passenger origins are highly concentrated, where the naive frequency method can reduce estimation error from 8.294 to 0.650. When comparing different estimation methods, the MV10 algorithm performs more stably during early stages with sparse data, while the naive frequency method achieves better accuracy after accumulating sufficient historical data. The case study using real data further validates the importance of historical data initialization: when using 28 days of data to train the estimator and adopting a strategy that dynamically switches estimation methods based on data volume, it outperforms the traditional non-estimation approach in objective values. These findings demonstrate the effectiveness and practical value of integrating real-time flow estimation into elevator dispatch systems.