Short-term prediction of parking availability in an open parking lot

3.1 Case study

The parking lot is located beside a shopping center in a mid-sized city in Sweden. A fraction of that parking lot is the region of interest (ROI) in this study, and it is highlighted by a green color line, as illustrated in Figure 1. A thermal camera is installed in the shopping center and is utilized to collect videos from the ROI. It was possible to collect data from an open parking lot without heavy installation and maintenance costs due to the use of the thermal camera. The thermal camera also avoided privacy concerns that would have been caused by the use of an optical camera [21]. The data for the entire parking lot could not be considered due to the limited field view of the camera. The ROI has four entrances, E1, E2, E3, and E4, and since it is located near the entrance of the shopping center, it is expected to have higher vehicle traffic [22]. The thermal camera views the vehicle from the side instead of the rear or front side of the vehicle, which hinders the visibility of the vehicle. Therefore, the parking occupancy data were collected manually.

Figure 1

Observed parking lot and the region of interest highlighted by the green line.

3.2 Data collection

There are previous studies that forecast parking occupancy information using months of historical data combined with real-time parking occupancy data. Such data can be collected for closed parking lots, where underground sensors can be utilized to collect long-term data. However, it is expensive to obtain such data for an open parking lot due to the lack of Parking Guidance and Information systems. Therefore, data collection was performed using videos from a thermal camera installed on the dome of the shopping center. Parking occupancy data were extracted manually using these videos. The number of parking spaces covered in this study is 47. The data for the forecast are collected over 21 days, between January 7 and January 27, 2020, as shown in Table 1. The first 19 days of data were used to train the model, while the remaining 2 days of data were used to evaluate the model. As mentioned in ref. [23], at least 5 days of hourly data are necessary to provide reasonable prediction values. Hence, the allocated training data size was sufficient to produce better prediction values. The parking spaces in an open parking lot are in higher demand and frequent changes within short intervals are expected. Therefore, the parking occupancy data were collected every 5 min. Since this article utilizes limited data for parking availability forecasting, it uses additional exogenous variables to improve the accuracy of forecasting. Exogenous variables as shown in Table 1 were utilized in this article. This consists of a total of six columns where the parking availability output variable is to be forecasted and the remaining are exogenous variables or predictors. Variables, such as level of traffic to the shopping center, day, holiday, and rain or snow, are exogenous as they are not dependent on the availability of parking spaces.

• Local time: This column provides the timestamp of the collected data, where the date and time are displayed. The data, as shown in Table 2, are captured between 9:00 and 20:00, with a frequency of 5 min. Each day represents 133 data points.

• Availability: This column is the parking spot availability represented in percentage, i.e., Availability = max { E j } 47 × 100 ; j ∈ w i , where E j is the number of empty parking spots out of 47 available spots at any point in time j , when the change in the empty parking spots occurs within a specific 5 min, the observation window w i , i = 1, 2, n. During peak periods, the availability becomes low, and during non-peak periods, it would be high. This is also the output or forecasted variable. The data on availability rate are illustrated through a time series plot in Figure 2.

• Level: This column is based on the visitor trend data from Google. However, the collected data are restricted to people using Google-based applications, such as Google Maps. Parking occupancy differs from each location of interest, such as restaurants, shopping centers, residential areas, or office buildings. Therefore, as mentioned in ref. [24], usage of parking profiles would facilitate improving the accuracy of the forecasting model. Popular times data are not available for certain periods, such as between 9:00 and 10:00, or between 19:00 and 20:00, due to few people visiting the shopping center. Since this only represents people using Google-based applications, these intervals are labeled as low. The periods where visitors were less than 60% are designated as medium. The periods near and around the peak, between 60 and 100%, are labeled as high. This variable is exogenous as visitor traffic may, or may not, affect parking availability. While visitor traffic is similar on weekdays, it varies on the weekend, as shown in Table 3.

• Day: This column represents days of the week. The use of this column facilitates forecasting parking availability based on the day of the week, and not on timestamp. As shown in Figure 3, the availability decreases by 12:00 during weekdays, while it increases by 13:00 during weekends. Availability during weekdays also sharply increases after 17:00 but gradually increases from 16:00 over the weekend.

• Holiday: This column represents public holidays in Sweden. Parking occupancy varies on holidays. Therefore, it is necessary to consider this variable to forecast parking availability during holidays. However, traffic variation during all the holidays is not the same. Traffic during Christmas is high when compared to other public holidays. This variation is not observed in this study due to the limited dataset.

• Rain or snow: This column describes if there is rain or snow. Weather conditions usually impact travel behavior. However, in a country like Sweden, we expect this impact to be minimal unless there is rain or snowstorm.

Figure 2

Illustration of the parking availability of the total dataset.

Figure 3

Parking availability during a weekday (January 18, 2020) and a weekend (January 19, 2020).

Approximately 2,400 data points were utilized to train the selected algorithms, while 400 data points were utilized to evaluate the predictive performance of the algorithms. Training and testing were performed with and without visitor trend data to evaluate its effect on the performance of the algorithms. Two sets of prediction methods were trained, where one set includes visitor trend data during training and testing, while the other set does not include visitor trend data. Each set consists of LSTM, SARIMAX, and Ensemble-based methods. The exogenous variables, such as holiday, rain or snow, and day of the week, are included as they were found to improve the accuracy of the algorithms in several previous studies [25,26,27,28]. As parking availability varies between weekdays and weekends, the test dataset consists of data from a weekday and a weekend. Prediction methods were evaluated multiple times, such as 09:00, 13:00, and 18:00 during these days. Dell workstation with Quadro P5200 GPU was utilized in this study along with Matlab and Anaconda platforms to prepare data and evaluate prediction methods.

3.3 Algorithms for short-term prediction

Multiple algorithms, such as LSTM, SARIMAX, and Ensemble-based methods, are evaluated in this article with and without visitor trend data. The exogenous variables mentioned in Section 3.1 are the predictors, while the parking availability information is the response of the prediction methods.

3.4 Evaluation metrics

Parking availability predictions were provided for 60 min. The predicted results were evaluated using the following metrics, where y i is the actual value at i the instance, while y ˆ i is the predicted value at instance i.

MAE:

Root mean squared error (RMSE):

4.1 Prediction with visitor trend data

Figure 4 illustrates prediction methods forecasting at multiple periods, namely 09:00, 13:00, and 18:00 during a weekend. The Ensemble-based method provided better short-term predictions than all other methods. As illustrated in Figure 4(b) and (c), the LSTM provided better predictions at 13:00 and 18:00, compared to SARIMAX. The sudden change in parking availability was learned better by Ensemble and the LSTM method. However, as shown in Figure 4(a) at 09:00, LSTM produced high errors. The SARIMAX method only produced better short-term predictions at 13:00, performing better short-term predictions when there is no sudden change in parking availability.

Figure 4

Prediction plot of methods with visitor trend data on the weekend. (a) Predictions with visitor trend data at 9 AM on weekend. (b) Predictions with visitor trend data at 1 PM on weekend. (c) Predictions with visitor trend data at 6 PM on weekend.

As shown in Table 4, the Ensemble-based method provided better predictions for multiple periods on a weekend with lower MAE and RMSE values. The total dataset consists of 3 weeks of data, where only 5 days representing a weekend were utilized for training the prediction methods. Despite the limited volume of data available for weekends, the Ensemble-based method produced better predictions. LSTM also performed better with fewer errors when compared to SARIMAX.

The overall performance of the three prediction methods was better for a weekday, as shown in Figure 5(a) and (c). However, when there is a sudden change in the rate of parking availability at 13:00 as shown in Figure 5(b), all three methods were unable to predict that change. Within 10 min, 10% of empty parking spaces were occupied. This random change in parking availability is expected for an open parking lot. However, due to limited data available for training and a lack of other potential exogenous variables, the sudden change in parking availability was not captured.

Figure 5

Prediction of parking availability with visitor trend data on weekday. (a) Predictions with visitor trend data at 9 AM on a weekday. (b) Predictions with visitor trend data at 1 PM on a weekday. (c) Parking predictions with visitor trend data at 6 PM on a weekday.

As shown in Table 5, SARIMAX produced lower MAE and RMSE values than other methods at 09:00 on a weekday. However, all three methods produced high errors at 13:00. LSTM produced fewer errors in predicting at 18:00. No method produced lower errors for all the mentioned periods. Ensemble-based method predicts values based on aggregation and there are not enough historical data to predict the change in parking availability during peak hours. LSTM would have produced better predictions with more training data.

4.2 Prediction without visitor trend data

In this section, the prediction methods were evaluated without the use of visitor trend data for multiple periods mentioned in Section 4.1.

Figure 6 illustrates the prediction values without the use of visitor trend data on a weekend. As illustrated in Figure 6(a), the Ensemble-based method produced better prediction results at 09:00, and LSTM and SARIMAX produce predictions with high errors. As shown in Figure 6(b), all three methods provided predictions with lower errors. However, at 18:00, as illustrated in Figure 6(c), SARIMAX predictions produced high errors. This performance is similar to the performance with visitor trend data. Therefore, there is no change in the performance of SARIMAX and Ensemble without the visitor trend data.

Figure 6

Prediction plot of methods without visitor trend data for the weekend. (a) Predictions without visitor trend data at 9 AM on weekend. (b) Predictions without visitor trend data at 1 PM on weekend. (c) Predictions without visitor trend data at 6 PM on weekend.

As shown in Table 6, the LSTM performed slightly better with visitor trend data when compared to without visitor trend data. The MAE and RMSE values were higher for LSTM without the use of visitor trend data. However, SARIMAX and Ensemble-based methods provided similar MAE and RMSE values with and without visitor trend data.

As illustrated in Figure 7, the predictions of the Ensemble-based method and SARIMAX were similar to visitor trend data. Predictions of LSTM were the only predictions affected when excluding the visitor trend data. Contrary to the performance of LSTM on a weekend, the predictions slightly improved over a weekday without the use of visitor trend data, as shown in Figure 7(a) and (c). The variations in parking availability over a weekday were better learned by the LSTM without the use of visitor trend data. As illustrated in Figure 7(c), at 13:00, the sudden change in parking availability was not captured by any prediction methods without visitor trend data as well. As shown in Figure 7(c), at 18:00, LSTM performed better, compared to other methods. SARIMAX still maintained lower parking availability at 18:00, which is the reason for larger errors.

Figure 7

Prediction plot of methods without visitor trend data for weekday. (a) Predictions without visitor trend data at 9 AM on a weekday. (b) Predictions without visitor trend data at 1 PM on a weekday. (c) Predictions without visitor trend data at 6 PM on a weekday.

As shown in Table 7, LSTM provided better predictions with lower MAE and RMSE values compared to the Ensemble-based method and SARIMAX. The LSTM predictions were slightly improved for multiple periods on a weekday when compared to predictions with visitor trend data. For methods, such as SARIMAX and the Ensemble-based method, predictions had no impact with or without visitor trend data.

Despite the less volume of training data, the Ensemble-based method and LSTM learned the patterns more efficiently similar to results observed in refs. [17,19]. Since there is a similarity in parking availability over weekends, the Ensemble-based method performed well, despite there being less volume of data. The Ensemble-based method utilizes multiple samples of the training dataset to generate aggregated values from decision trees and random forest, which is the reason for the improved prediction, but when there are quick changes in parking availability, which are not based on previous data, then the prediction methods would produce high errors in their predictions. However, this can be addressed by adding relevant exogenous or endogenous variables that affect this behavior. The LSTM and SARIMAX methods’ predictions are based on previous values. However, due to the usage of a memory cell in LSTM, it understood the time series data efficiently compared to SARIMAX. SARIMAX produced better prediction in ref. [14]. However, in this study, due to less volume of data, SARIMAX produced high errors when there are sudden changes in the rate of parking availability. By increasing the volume of the dataset, the performance of the methods can further be improved. Another reason for the poor performance of SARIMAX is due to the lack of parking time information.