Big news: Tryolabs joins Qubika. Same teams, broader set of combined capabilities. Read now ->

Case study

Occupancy prediction in smart parking lots

The developed models were integrated within the client's in-house business analytics platform, allowing them to seamlessly analyze the outcome of the predictive models.

Industry: Transportation & logistics
Client
Solution: Demand forecasting

Occupancy prediction in smart parking lots

Demand forecasting up to 1 day in advance

The Luxembourg National Railway Company or SNCFL is one of the leading economic players in the Grand Duchy. With 22 companies and more than 4.800 employees, the CFL Group ranks among Luxembourg's largest employers. Willing to provide a user-friendly transportation service, CFL has been working with us to predict the occupancy rate of two of their parking lots adjacent to train stations to make it easier for commuters to park at their facilities.

Challenge

The client required a custom machine learning model that could predict the occupancy rate of parking lots adjacent to train stations. Having that forecast could encourage users to take the train more often, without the fear of driving to the train station to find out that no parking spots were available.

CFL installed smart sensors in two of their parking lots in 2020, and they have been collecting and centralizing the data ever since in Pentaho: a business intelligence software hosted in-house.

Since CFL already invested in the platform and the necessary infrastructure to host it, the developed solution had to be able to integrate into their existing workflow, allowing their employees to continue using the same tools that they were previously using to explore and process the new data.

Solution

The client envisioned that the predictive model for parking occupancy could be used in two cases. First, to provide passengers about to drive to the train station an estimate of how full the parking lot will be when they arrive. They also wanted to offer their customers the possibility to plan their trip for the next morning the night before. For the first use case, a short-term model capable of forecasting up to 4 hours ahead was built, while a long-term model was built for the second use case, as day-ahead predictions were required.

Our team implemented the models while satisfying an important constraint: since this was a proof-of-concept problem to show that a custom-made solution could be deployed within their infrastructure, the only dataset available consisted of previous measurements of occupancy prediction for the two parking lots.

1 week

for first prototype running in production

<5%

error in predictions

Discover our approach

Our solution consisted of two machine learning models for time-series prediction deployed on the client infrastructure. These models are periodically updating the predictions to provide real-time information to passengers.

The data for this project was obtained from a third-party company that installed sensors in two parking lots in Luxembourg. From these sensors, our client calculated the number of occupied spots in each parking area.

At the moment of starting the project, the client had identified that some sensors were producing noisy measurements. This was mitigated by re-sampling the data to a 5-minute granularity, avoiding some errors.

Even though the problem belongs to the traditional time-series forecasting literature, we used a modern approach consisting of extreme gradient boosting trees, as they facilitate the extension of the model with more complex features in later stages.

Since regression trees do not natively understand temporal data, we extracted several features from the date of the measurement that could be used by the model to distinguish between hours and days. Among them, we extracted minutes, hours, days and weeks (which we embedded into a 2D space to preserve periodicity), but also lag features (the occupancy rate at an earlier point in time).

One of the advantages of deploying the models within the client's infrastructure, is that it enabled them to reuse their existing tools to analyze the newly created data. In particular, a dashboard was created to showcase how a passenger might preview the prediction but that can also be used to monitor the performance of the model.

The client was already invested in using PowerBI to create dashboards which had direct access to their databases. By saving the predictions to those same databases, CFL could benefit from their existing knowledge and infrastructure to observe and use the output of the model in real time.

Discover how the world’s leading enterprises partner with Tryolabs for building data-driven solutions that boost their KPIs.