CASE STUDY

Dynamic Demand Forecasting System for E-bikes as Last Mile Connectivity

Industry

Transport & Logistics

Work Done

Challenge

Client has a fully automated public bike sharing system that spans across various cities in India, including Ranchi, Surat, Kolkata, and Prayagraj. The client was looking for a dynamic demand forecasting system to predict customer demand and reduce costs and overhead in a more accurate manner.

The prediction of needs is critical due to the distribution of the limited resources (bikes and empty slots to place the bikes) and the management of the bike demand sharing system.

The challenge while implementing the system was defining external data points as each city offers different dynamics, seasons, holidays, and cultural distribution in India.

The solution

We initiated the project by understanding the application and collecting various internal data points required for the success of the application. We tried to analyse external factors such as day, time of the day, weekday/weekend, holiday, season, weather, etc. as external factors to observe an impact on industry in a systematic manner.

This enabled us to provide models that took advantage of the uneven distribution of bikes among stations, which is frequently unbalanced and causes numerous problems in daily operation.

Predicting bike user count throughout the day contributes to providing a prominent solution for maintaining required bike availability at stations before a requirement arises. The solution puts forward a bike demand prediction model based on a multivariate supervised LSTM deep learning algorithm that considers external weather parameters with the least amount of training data. The prediction model’s components include user-centric parameters such as city name, station name, and the number of forecasting days. Once the user hits the user-centric parameters, the model provides features such as forecasting dates, day parts, and bike user counts.

To conclude our solution, we created analytics to assist management decision-making with all relevant data demonstrating the predicted bike demand results, overheads, and a suggestive solution to reduce overheads.

The top-most graph depicts the predicted outcome for each part of the day based on the weather parameters. The bottom graph demonstrates the average error in each part of the day corresponding to a weather parameter. The right-side graph demonstrates the effect of actual and predicted demand on the entire day and its corresponding error.

Technologies used

Results

A dynamic demand prediction system and analytics solution was provided to the client in a span of 8 months that assisted the client in reducing overhead costs by 8.6%, and the movement of bikes from one place to another based on demand further reduced the cost by 7%. The demand prediction also helped them identify where they were doing good and where they had scope for improvement, which resulted in better marketing and budgetary allocation that resulted in overall 4.76% YOY growth in 2022.