- There are anomalies and a lack of data in the raw data corresponding to the train station garage: 2016/08/06 – 2016/08/31, 2016/12/06 – 2017/25/06, 2018/04/11 – 2018/05/18, 2020/09/03 – 2020/09/11, 2020/09/20 – 2020/10/16 and 2021/06/18 – 2021/07/21. These anomalies in data need to be explored to select the appropriate value for model training and validation.
- Different data range selections had a major effect on model metrics, surpassing even the effect of the hyperparameters tuning.
- The optimal parameters of Prophet algorithm are: changepoint prior scale = 0.001, seasonality prior scale = 10, yearly seasonality = True, weekly seasonality = True, daily seasonality = True
- A different seasonal behavior was identified from the second half of 2020 to the first quarter of 2021, leading to a significant increment in minimum daily occupation. A decrease followed this behavior in the occupation rate of the train garage. The monthly and seasonal variations observed in 2020 and 2021 influenced the model performance on prediction, especially to April 15th, 2022, between 12:00 and 13:00, considering that the available data finish in 2021/12/31.
- The appropriate selection of training and validation data represents the most important parameter to enhance the model performance, followed by time shift and hyperparameter tuning. For this reason, the ensemble method could be an affordable solution to increase prediction accuracy.
Traffic volume in urban areas is continuously growing due to massive urbanization. The increasing traffic makes urban life more congested and polluted, increasing energy consumption, airborne diseases, and global warming. According to NASA analysis, 2020 has been recorded as the hottest year.
In this scenario, ensuring efficient traffic movement in smart cities is an important goal, minimizing the negative environmental and public health impact. Nowadays, a common problem of huge cities is based on searching for a free parking space, which generates an increase in traffic congestion, fuel consumption, and greenhouse gas pollution in urban areas.
A potential solution to mitigating parking searches is accurately predicting parking availability, which reduces drivers waiting time and traffic congestion while looking for a free parking slot.
In this project, the following tasks are performed:
- Data analysis
- Parking garages comparison
- External data sources incorporation: Public holidays and weather forecasts.
- Data processing and Feature engineering
- Building a predictive model of parking availability: Prophet (Univariate/ Multivariate).
- Hyperparameter tuning.
The following assumptions were assumed in the development of the predictive model:
- The parking capacity of both garages remains constant in the all-time data range.
- Specific social events on historical data are negligible.
