AI predicts parking availability by using weather, traffic speed, and meter data

We’ve all been there: You drive miles to a venue only to discover that, to your dismay, every parking space is fully occupied. Apps like Google Maps, which can predict busyness based on historical data, can help to a degree, but what if you’re in need of a more adaptable solution? Enter research by scientists at Carnegie Mellon University, who describe in a newly published paper on the preprint server Arxiv.org an AI system for predicting parking occupancy in real time.

Rather than collect data from parking sensors, which the study’s coauthors contend are susceptible to failure and error, they draw on parking meter transactions to first estimate parking availability before using additional data for prediction. An estimated 95 percent of on-street paid parking is managed by meters, making their model more generalizable than sensor-dependent systems.

“In this study, we adopt the data-driven approach by incorporating multiple traffic-related sources, in terms of both real-time and historical data, including parking occupancy, traffic conditions, road characteristics, weather and network topology,” the researchers wrote. “It ultimately predicts (or forecasts) short-term parking occupancy via a deep neural network method.”

The team used a graph convolutional neural network — an algorithm that operates on nodes, edges, properties, and other graph structures — to model the statistical relationship among parking locations, traffic flow, parking demand, road links, and parking blocks. Together with a recurrent neural network with long-short term memory (LSTM) — a type of AI algorithm capable of learning long-term dependencies — and a multi-layer decoder, the system extracted parking information from traffic-related data sources (such as parking meter transactions, traffic speed, and weather conditions) and output occupancy forecasts.

Above: A heatmap demonstrating improvements in the model by incorporating weather conditions.

The researchers trained it on data sourced from the Pittsburgh downtown area, which they note has 97 on-street parking meters across 39 street blocks. Historical parking stats came from the Pittsburgh Parking Authority, while connected car company Inrix’s Traffic Message Channel and WeatherUnderground’s API supplied traffic speed data and hourly weather reports, respectively.

In tests, the model outperformed others’ baseline methods when predicting parking occupancies 30 minutes in advance, the researchers say. They credit the weather and traffic speed data for the AI system’s superior performance — particularly the weather data, which boosted prediction accuracy in recreational areas.

“In general, lower prediction errors are received on blocks with larger parking capacities,” the paper’s coauthors wrote. “It is no surprise as higher parking capacities usually result in lower variances in occupancy rate the model performs better on business districts … [P]arking demand in business districts usually has strong daily patterns, and is more resilient to impacts from unusual scenarios such as hazardous weather and special events, which has made prediction more efficient.”

They leave to future work a model that incorporates additional traffic-related data, including traffic counts, road closure, incidents, and events.