دانلود مقاله مدل شبکه عصبی مبتنی بر هوش مصنوعی و پیش‌بینی تقاضای مسافران با داده های کارت هوشمند
چکیده
1. مقدمه
2. منبع داده و روش شناسی
3. توسعه مدل شبکه عصبی مصنوعی
4. دقت پیش بینی کوتاه مدت
5. ارزیابی تطبیقی با سایر مطالعات موجود
6. نتیجه گیری و مسیرهای تحقیقات آتی
منابع
Abstract
1. Introduction
2. Data source and methodology
3. Artificial neural network model development
4. Short term prediction accuracies
5. Comparative evaluation with other studies in the literature
6. Conclusions and directions for future research
Ethics clearance
Acknowledgements of Research Funding
References
چکیده
پیش‌بینی کوتاه‌مدت دقیق تقاضای حمل‌ونقل عمومی برای عملیات حمل‌ونقل عمومی بر اساس تقاضا ضروری است. دانستن اینکه کجا و چه زمانی تقاضاهای آینده برای سفر مورد انتظار است، به اپراتورها اجازه می دهد تا به سرعت جدول زمانی را تنظیم کنند، که به بهبود کیفیت و قابلیت اطمینان خدمات و جذب مسافران بیشتر به حمل و نقل عمومی کمک می کند. این مطالعه با توسعه مدل‌های یادگیری عمیق مبتنی بر هوش مصنوعی برای پیش‌بینی تقاضای مسافران اتوبوس بر اساس داده‌های حمایتی واقعی به‌دست‌آمده از سیستم فروش بلیط کارت هوشمند در ملبورن، به این نیاز می‌پردازد. این مدل‌ها که ویژگی‌های زمانی تقاضای سفر برای برخی از سنگین‌ترین مسیرهای اتوبوس در ملبورن را در نظر می‌گیرند، با استفاده از داده‌های دنیای واقعی از 18 مسیر اتوبوس و 1781 ایستگاه اتوبوس توسعه یافته‌اند. مدل‌های یادگیری عمیق LSTM و BiLSTM با استفاده از مجموعه داده‌های مشابه با پنج مدل یادگیری عمیق مرسوم ارزیابی و مقایسه شدند. مقایسه دسکتاپ نیز با تعدادی از مدل‌های پیش‌بینی تقاضای تثبیت‌شده که در ادبیات دهه گذشته گزارش شده‌اند، انجام شد. نتایج ارزیابی مقایسه ای نشان داد که مدل های BiLSTM از سایر مدل های آزمایش شده بهتر عمل کرده و قادر به پیش بینی تقاضای مسافران با دقت بیش از 90 درصد هستند.
توجه! این متن ترجمه ماشینی بوده و توسط مترجمین ای ترجمه، ترجمه نشده است.
Abstract
Accurate short-term forecasting of public transport demand is essential for the operation of on-demand public transport. Knowing where and when future demands for travel are expected allows operators to adjust timetables quickly, which helps improve service quality and reliability and attract more passengers to public transport. This study addresses this need by developing AI-based deep learning models for prediction of bus passenger demands based on actual patronage data obtained from the smart-card ticketing system in Melbourne. The models, which consider the temporal characteristics of travel demand for some of the heaviest bus routes in Melbourne, were developed using real-world data from 18 bus routes and 1,781 bus stops. LSTM and BiLSTM deep learning models were evaluated and compared with five conventional deep learning models using the same data set. A desktop comparison was also undertaken against a number of established demand forecasting models that have been reported in the literature over the past decade. The comparative evaluation results showed that BiLSTM models outperformed other models tested and was able to predict passenger demands with over 90% accuracy.
Introduction
A well-developed urban public transport system, especially bus transport, can reduce congestion and emissions and decrease the use of private vehicles (Li, Cao, et al., 2020). On-demand public transport, in particular, is seen to have the potential to improve operations further and enhance customer satisfaction. However, this type of service requires that short-term forecasts of future demands for bus services are known in advance (Liang et al., 2019; Liyanage & Dia, 2020; Liyanage et al., 2019; Smith et al., 2002; Zhou et al., 2013). Accurate prediction of future demands also helps operators to pre-allocate constrained resources such as vehicles and drivers to meet passenger demands and provide quality and reliable services with minimum waiting times. It also allows operators to optimize bus fleet management to minimise operational costs (Ma et al., 2014; Tirachini et al., 2013). Demand prediction is an integral part of business and commerce operations and helps decision makers to reduce the uncertainties of future operations. In public transport operations, the passenger service business models are highly dependent on accurate estimation of future passenger demands. Starting from route design and network planning, through to scheduling of vehicles with optimised seating capacity to meet operators' and users' objectives, and pricing each passenger vehicle on a network route, service operations in every planning horizon is dependent in one way or another on accurate estimation of future demands (Banerjee et al., 2020; Lu et al., 2021).
Conclusions and directions for future research
This paper presented robust deep learning models that were developed for short-term temporal predictions of passenger demand using real-world data. The data was obtained from the MyKi smart-card fare payment system in Melbourne. Deep learning models were constructed representing 15-min, 30-min, and 60-min. These models were developed using BiLSTM and LSTM deep learning methodologies based on one month of data comprising 27,823 data points for the 15-min model, 27,360 data points for the 30-min model, and 26,561 data points for the 60-min model. The findings of this study showed that both the BiLSTM and LSTM architectures provided the highest predictive intelligence accuracy of over 90% for short-term predictions of passenger demands for 15-min, 30-min and 60-min time horizons.
The main limitation of this study was limited access to smart-card data which in our case was constrained to one month of data between 1–27 May 2018. This was the only data made available to the researchers. Access to larger data sets covering more months and years as well as more routes can help improve the accuracy and reliability of the models even further. Furthermore, for this analysis, aggregated passenger demand for inbound and outbound direction for each route was necessary because the data did not have sufficient observations to enable separate analyses for inbound and outbound directions. It is noted here, however, that the direction of the service will not have an impact on model selection and performance. The key factor influencing performance, in either inbound or outbound directions, is the availability of quality representative data that can be used for model training and testing. We aim to address this limitation in future studies through analysis of large data sets that can provide accurate and sufficient observations in both directions of travel.