Results of statistical models of parking meter payment data were compared against data collected by sensors in the SFpark pilot program to compare accuracy
Between 2011 and 2013 the San Francisco Municipal Transportation Agency (SFMTA) implemented the SFpark pilot project, which used in-ground sensors and new parking meters to collect real-time data so parking prices could be adjusted based on occupancy.
The sensors were costly to implement and had a short operational life. However, San Francisco upgraded all of its parking meters to collecting payment data as part of SFpark. Thus, while occupancy data is no longer available, payment data is still available. The SFMTA analyzed sensor and meter data collected during the SFpark pilot and developed a sensor-independent rate adjustment (SIRA) model estimating parking occupancy using meter payment data, enabling demand-responsive pricing policies without large-scale sensor installations. This report describes how the SIRA model works and compares its results with data from the SFpark pilot period when parking spaces were equipped with sensors.
Three models were developed to estimate parking occupancy using meter payment data focused on two metrics: parking occupancy rate and meter payment rate. Occupancy and payment rates could be analyzed at the block time band level based on data collected in the SFpark pilot, where rates could be adjusted based on three criteria: block, generally defined as both sides of a street between major intersections; time of day, divided into 3 time bands: 9am to noon, noon to 3pm, and 3pm to 6pm, and; week or weekend day.
A simple linear regression, where occupancy rate was the dependent variable and payment rate the independent variable, found a strong, statistically significant, positive relationship. More meters are generally paid when more spaces are occupied. Conversely, when fewer spaces are occupied, fewer meters are paid. Based on the results of this model and analysis of the residuals, two additional models, a multiple linear and log-log model, were tested with five additional variables: four dummy variables representing unique parking management districts and one to account for weekday versus weekend non-payment trends. If no dummy variables are applied, the model estimates neighborhood commercial areas on weekdays. Model evaluation criteria: R-squared, accuracy of the rate adjustment outcome compared to the observed SFpark data, and distribution of error.
Each model estimates occupancy accurately 67-69% of the time and within $0.25 29-30% of the time. The basic model was least accurate and log-log model with predictor variables most accurate. The multiple linear had the highest r-squared value. The multiple linear regression was selected for implementation because it had the highest r-squared and, when it was not accurate, it tended to set the price too low, favoring users.
Since the model is mechanistic, an error in occupancy prediction during one adjustment is corrected in the next adjustment. This is a strength of this design. While the SFMTA cannot estimate what the right price is, the agency can adjust prices in response to demand. As long as the SFMTA continues to regularly adjust on-street meter rates, it will get prices right in the long run.