The global electric car sales continued to exceed the expectations climbing to over 3 millions and reaching a market share of over 4%. However, uncertainty of generation caused by higher penetration of renewable energies and the advent of Electrical Vehicles (EV) with their additional electricity demand could cause strains to the power system, both at distribution and transmission levels. The present work fits this context in supporting charging optimization for EV in parking premises assuming a incumbent high penetration of EVs in the system. We propose a methodology to predict an estimation of the parking duration in shared parking premises. The final objective is estimating the energy requirement of a specific parking lot, evaluate optimal EVs charging schedule and integrate the scheduling into a smart controller. We formalize the prediction problem as a supervised machine learning task to predict the duration of the parking event before the car leaves the slot. We test the proposed approach in a combination of datasets from 2 diferent campus facilities in Italy and Brazil. The overall results of the models shows an higher accuracy compared to a statistical analysis based on frequency, indicating a viable route for the development of accurate predictors for sharing parking premises energy management systems.
an estimation of the parking duration in shared parking premises. This is essential for estimating the energy reThe advent of Electrical Vehicles (EV) are in increasing quirement of a specific parking lot, evaluate optimal EVs spreading in our society. According to MCkinsley re- charging schedule and integrate the scheduling into a port1 in our society EV sales rose 65 percent from 2017 smart controller. to 2018 and Europe has seen the strongest growth in The specific behaviour of parking lots of campuses refEVs. According to the report "The European Union’s ereed to EV charge is peculiar since it substantially difers new emissions standard—95 grams of carbon dioxide per from the general parking lots available in the streets. In kilometer for passenger cars—could also boost EV sales campus-like facilities (Universities, large industries, etc) because it stipulates that 95 percent of the fleet must we can observe regular patterns of parking behaviour meet this standard in 2020 and 100 percent in 2021". A that mainly include staf working hours besides a part of race for larger batteries among manufacturer is leading other visitor [1]. This can be an advantage when trying the current EV technology, and going forward it appears to predict general behavioral patterns of parking habits that as batteries technology improve, they are going to and thus reach an optimal recharge plan for EVs. replace motor fuel vehicles. The concerns as we move to The current work can become part of an overall deEVs is that, firstly, there will not be enough charge points sign of a smart charging energy management system to to meet consumer demand and, secondly, this additional optimally integrate the distributed energy systems and load on the electricity grid will cause partial and total EVs into the power grid by developing a parking predicfailure of specific electrical plant due to overloading. The tion module to estimate the vehicles’ parking time using present work fits this context supporting optimization for machine learning algorithms.
EV charging and assuming a incumbent high penetration Given this context, the specific objective of this work is of EVs in the system. We propose an approach to predict to predict the duration of each parking event in a campuslike parking lot, where the parking event is the moment of Proceedings of the Workshop on Big Mobility Data Analytics (BMDA) the actual parking action of a car in a slot. In other words, co-located with EDBT/ICDT 2023 Joint Conference (March 28-31, at the moment the car is parked in a slot, we predict how 2023), Ioannina, Greece long this car will stay parked so that the change can be $ viniciusmonteirodelira@isti.cnr.it (V. M. d. Lira); optimized. This will allow the energy management syslfoabreiannzoo..pgaabllroineelltit@o @istmi.cun.ire.it(F(.LP.aGllaobnreietltloi));; chiara.renso@isti.cnr.it tem to decide when to start the actual charge based on (C. Renso) the prediction. For example, assuming several cars arrive CPWrEooUrckReshdoinpgs IhStpN:/c1e6u1r3-w-0s.o7r3g ©CCo2Em02mU2oCRnospLWyicreiognhsrtekfAostrthrtihboiusptpioanpPe4rr.0obIynctietesrenaaduttiihononragsl.s(CUCs(eCBpYEer4mU.0i)tR.te-d WundSe.roCrregat)ive aatn9d AthMer, eiffowree dwoen’sttkarntocwhahrogwinmguimchmtheedicaatrelsyt,aythpisarwkeildl 1https://www.mckinsey.com/industries/automotive-and- cause a peak of electricity demand. On the contrary, if we causssehmiobnlsy-/ao-ugrlo-ibnasli-gphlutsn/gmec-ikni-nesve-ysa-elelesc#tric-vehicle-index-europe- predict that a given car stay parked e.g. 8 hours, then we can delay the charge of this car to a later moment trying therefore predict the subsequent parking availability comto flatten the peaks. We formalize the prediction prob- bining a matching-based allocation strategy to assign lem as a supervised machine learning task that, given a users to selected parking spaces. parking event at a given time, tries to predict the dura- Deep learning to predict parking occupancy is protion of the parking event. We structure our experiments posed by many papers in the literature. Paper [5] adopts inspired by the research questions: RQ: How accurately a deep learning model for predicting block-level parka supervised machine learning approach can predict the ing occupancy 30 min in advance in paper. The model duration of a parking event in a campus-like parking lot? takes multi-source data as input, e.g., parking, trafic and We experiment diferent algorithms and features combi- weather. nation into 5 datasets from 2 diferent campus facilities Paper [6] proposes a Convolutional Neural Network in Italy and Brazil. We show that using both contextual (CNN) model for block-level parking occupancy predicand time of the day features, the overall results of the tion extracting spatial relations of trafic flow combined models shows an higher accuracy compared to a statisti- with a LSTM autoencoder to capture temporal correlacal analysis based on frequency, indicating a viable route tions. Clustering is also considered in the Clustering for the development of accurate predictors for sharing Augmented Learning Method (CALM) to learn deep feaparking premises energy management systems. ture representations of spatio-temporal data obtained
Structure of the paper follows. Section 2 discusses using the proposed embedding. some related works an how this approach diferentiates A deep learning approach is also proposed by [7], from them. Section 3 introduces the problem definition called Du-Parking. This approach models temporal closeand clarifies the prediction problem formulation. Section ness, period and current general influence employing 4 reports the details of the experimental evaluation. Fi- long short-term memory (LSTM) to model the temporal nally, Section 5 draws the conclusions and envisage some closeness and period. This approach learns to dynamifuture works. cally aggregate the output to estimate the final parking availability of given parking lot.
Compared to the above approaches the novelty of our 2. Related Works method lies in the fact that we do not aim at predicting the next free slot neither to suggest the driver where to park. On the contrary, we predict how long the car will stay parked in a given slot and this prediction is computed at the specific time the car start the parking, which is a diferent problem. While the first problem requires to train a learning system to predict which slots - in a given area at a given time - will be free, this problem, given a specific slot and a specific starting time of the parking of a vehicle, predicts how long the car will stay parked.
The energy management system can therefore allocate the energy to the parking slot charging station based on this prediction: charging later in time vehicles which are predicted to stay longer and accelerate the charging for cars which are predicted to stay shorter.
Most of the recent works in the literature focuses on predicting which parking lot will be free at the arrival of the car. This is motivated by the challenge of finding a parking space in urban areas. It has been reported2 that 30% of trafic congestion is caused by travelling for ifnding parking spaces, bringing unnecessary energy consumption and environmental pollution. Works in this area include the of-street (parking slots in private areas) and in-street variants (slots in the streets). A pioneering paper by [2] proposes the real-time availability forecast algorithm to predict parking facility availability in real time using combined current (on-line) and historical information. This work uses an algorithm operating with mixed real and simulated information.
A recent paper by Zeng et al ([3]) proposes a hybrid model that stacks gated recurrent unit (GRU) and longshort term memory (LSTM). The GRU-LSTM model combines LSTM’s advantage in prediction accuracy and GRU’s advantage in prediction eficiency. Furthermore, similar to us, it uses multi factors, including occupancy, weather conditions and holiday, as input to predict parking availability.
In paper [4] authors develop a prediction model based on Naive Bayes and machine learning methods like decision tree, random forest, and regression analysis for building the prediction model of parking occupancy and
The objective of our approach is to exploit historical data on parking usage and additional contextual data like weather conditions and parking lot occupancy levels, to predict the duration of a parking slot occupancy. Diferently from many state of the art approaches that want to predict if a giving parking lot will be free in a next period of time [2, 4], here we focus on the prediction of the temporal duration of the occupancy of a car in a slot.
We recall that our approach, to be suitably integrated with an Energy Management System, focuses on specific 2https://www.accessmagazine.org/spring-2011/free-parkingfree-markets/ parking context that we call of shared premises (e.g. park- to define a function (_, ) = where the class ing lots of universities, workplaces, supermarkets, etc), represents a temporal interval such that ⊆ . not focusing on fee-based street parking. Overall, our approach can be applied to any parking environment where there is a tendency for the car to stay parked a minimum amount of time and where the electrical charge system of the parking lots can be integrated into the controller of the Energy Management System. We also recall that our approach is driver-profile agnostic due to privacy reasons.
It is worth noticing that the parking behaviour in a campus-like facility reflects a diferent parking behaviour compared to fare-based streets parking lots since here the duration is expected to be longer than on street parking.Furthermore, the premises have usually a controlled access and the energy management system can therefore optimise the electricity supply based on the parking occupancy, while is not always true in fare-based street parking.
Given a parking area, a car parking event represents an event where a driver parks at a given timestamp in one of the available slots. The vehicle stays parked for a certain temporal duration until it leaves the slot. It is assumed that the vehicle can be charged while parked.
The charging time can start as soon as the car arrives, or can start later on, or again, can start, interrupt and start again.
ordinal categories. An ordinal variable is a categorical variable, where there is a clear ordering of the categories.
For example, our variable could assume ordinal categories like: short, medium or long duration. In the next section, we introduce the details of the Machine Learning (ML) approach to solve the Car Parking Duration Prediction Problem.
We propose to use supervised machine learning approaches to predict the parking duration based on an historical dataset of car parking events and contextual features.
The learning task is based on a three types of features: single event-related, spatial and contextual features. The event-related features represent the features that we can extract directly from the sets of parking events like the time of the parking event or the weather conditions. The spatial features are based on the location of the parking slots inside the car parking area, while the contextual features representing the occupancy of the diferent zones of the parking area. We explain how we have extracted some spatial and contextual features that are used in our predictive models. All these features are combined to feed the proposed supervised machine learning algorithms (Section 4).
Recalling Definition 2, a car parking event is defined Definition 1 (Parking Slot). Given a parking area in shared premises, we define a parking slot as a tuple =< , , >, where represents the parking slot identifier, and and represent its spatial coordinates in the parking area. A parking slot is the actual place where drivers park their cars. In our application scenario, each slot can be equipped with a charge station where the car can be recharged. The set of all the parking slots form the parking area. by the tuple =< , , > and given and we want to predict the temporal duration . From the timestamp , we derive three features: the day of week , hour of the day ℎ, and the minutes rounded to 5 minutes. The motivation of these temporal features is to enable the predictive model to learn the correlation between the time when the car parks and the relative parking temporal duration. We also include in this category of features the weather condition at the moment Definition 2 (Car Parking Event). We define a car park- of the car parking event starts, , using this as extra ing event as a tuple =< , , >, where information to feed the predictive models. represents the parking slot identifier where the car is parked, Many studies have been made toward the understand represents the timestamp indicating when a car has ing of parking behavior and the mechanism of people’s started the parking and is the temporal duration of the parking decisions [8, 9, 10, 11]. It has been observed that car park until it leaves the slot. some spatial aspects can bias the occupancy of a parking lot and choose the parking areas close to the destination We want to predict the parking duration of a car but also investigated the impact of the trees in parking parking event , given a and the parking event lots. starting time . This prediction is modelled as a Motivated by these aspects, our approach focus on classification problem where the objective is to assign, the spatial distribution of the parking slots. For this for each car parking event , a class representing the reason, we split the whole parking lot into smaller areas predicted duration interval. More formally, we have the using diferent clustering approaches. Then, we include following definition of the problem. these spatial features in our predictive models to learn Definition 3 (Parking Duration Prediction Problem). if a parking area can correlate with the slot occupancy Given a parking event where it is known the slot identifier duration. _ and the start time but not duration , we want More formally, given a set of parking slots (0, ..., ) ∈ 4. Experimental evaluation , we use the spatial coordinates of each to create the spatial clusters (1, 2, .., ) of parking slots, where ≤ . We have used two clustering algorithms for this In this section we evaluated the proposed approach for task: DBScan ([12]) and K-Means ([13]). Thus, when predicting the parking duration by exploiting historical training our predictive models over the the dataset of parking data. historical car parking events, we add as input feature a The research question driving our experiments is the representation of the cluster where the parking slot following: belongs to. RQ: How accurately a supervised machine learning ap
Another aspect that we investigate for the parking proach can predict the duration of a parking event in a duration prediction is the context. In our case the context campus-like parking lot? is represented by the status of occupancy of the slots in This research question guides our first experiments. the spatial clusters and the relationship of this occupancy Here, we compare the performance results of our mawith the duration of a given parking event. chine learning based approach against several baselines.
Specifically, we want to discover if the occupancy sta- We also investigate diferent machine learning approaches tus of an area (e.g 100%, means totally full, while 0% to tackle this problem as a supervised task: Classification, totally empty) where a driver parks, has a correlation Ordinal Regression, and Regression. We use diferent with the parking duration. features including the description of the occupancy of
The contextual features therefore represent the sta- the parking lot at the time the parking event starts and tus of occupancy of the diferent areas (i.e. the spatial the spatial distribution of the parking areas. clusters) of the parking. In other words, to predict the duration of a given car parking event , we also con- 4.1. Experimental Setup sider as input feature the level of occupancy of the spatial clusters (1, 2, .., ) at the time of the event. Datasets. We selected two public datasets of parking
More formally, we define a function called occupancy in campus-like parking lots: PKlot [14] and that given a and the timestamp of a parking CNRPark [15]. Both datasets contain the occupancy inevent , creates a vector (1 , 2 , .., ) by comput- formation detected by video cameras for each slots of ing the occupancy level of all parking spatial clusters parking areas of two academic institutions: the research (1, 2, .., ) at time . The occupancy status area of the National Research Council of Pisa3, in Italy of each spatial cluster is basically the ratio between and the parking area of the two Brazilian universities. the number of occupied slots at the timestamp and In both cases the whole parking lot is split in diferent the number of slots in that cluster. parking areas with a variable number of parking slots.
In Figure 1, we illustrate the spatial and contextual In both datasets, a car parking event occurs when a car features. At the top, we can see images from three dif- parks in a parking slot of the area. In this case, the event ferent car parking lots. In the middle, the dots represent starts at the timestamp of the frame that detects a car in the pixel coordinate of the parking slots. The colors of the slot. The car parking event ends at the timestamp each dots is a representation of the spatial features which of the frame showing: (1) an empty parking slot, or (2) indicates the cluster (sub-area) of each parking slot. a diferent car parked in the same slot. The duration of
In the bottom we illustrate the contextual occupancy the parking event is then computed as the diference of features. The color represents the spatial cluster of each the timestamps of the two image frames, the start and slot. At the bottom, we have the occupancy status in the end. The CNRPark dataset contains images collected percent of the spatial clusters 1, 2, 3, and 4 at a given from November 2015 to February 2016 for a total of 23 time . monitored days. The parking lot has been monitored by
To summarize, in this section we introduced three new 9 cameras covering the parking area of the Pisa National categories of features: event-related, spatial and contex- Research Council south parking area. The meteorologtual. The event-related features are extracted from the ical conditions at the moment of the frame capture has parking event, the spatial features are computed by us- also been collected. This dataset contains a total of 4081 ing cluster techniques over the spatial distribution of the frames and 144,965 photos. In Table 1 we depict an exparking slots while the contextual features are obtained cerpt of the CNRPark raw data. by computing the occupancy status of the spatial clusters. The change of cars in the monitored slots is also de
The overall idea is to investigate how to train the pre- tected and these statistics are reported in Table 2 for the dictive models using diferent information that might CNR dataset. We see the day of the week, the number of have a predictive power on the parking duration. In the changed cars (the end of a parking event) and the total next section we detail the experimental setting and re- number of days for which we have actual images. We sults on exploiting these features in a machine learning task for predicting the parking duration of a given event. 3http://www.area.pi.cnr.it also notice, as expected, how the number of car changes during the week end is very low.
The PKlot dataset contains the occupancy information for each slot of the parking areas of two academic institutions: (1) the Federal University of Parana (UFPR) and (2) the Pontifical Catholic University of Parana (PUCPR), both located in Curitiba, Brazil. The dataset includes a total of three diferent parking lots represented by PUCPR, UFPR04, and UFPR05. The occupancy information is detected by a number of cameras taking images of the parking slots and detecting the change of the car or the slot becoming empty. This dataset contains 12.417 images captured in three diferent parking areas with diferent weather conditions for a total of 168 slots in the period between 11 September 2012 and 16 April 2013. Specifically, dataset PUCPR has 100 parking slots, UFPR04 has 28 and UFPR05 has 45 slots. PKLot is larger than CNRPark and contains images spanning across months.
Data Cleaning. A detailed analysis on the image frames reveals the presence of missing data for some hour of the day (e.g. due to a broken device or during night hours due to the lack of infrared vision). When a parking event is starting or ending during the missing temporal interval, it has been flagged as partial and filtered out to avoid training inconsistencies. Additionally, we filter out the days when the number of detected slots is lower than 50% of the total slots. In total, after the data cleaning CNRPark counted 3552 parking events, PUCPR 4291 parking events, UFPR04 1204 parking events and UFPR04 2148 parking events.
Target Classes. We have considered the following 3 classes for the predictive variable (i.e. the car parking event duration) with discrete values in minutes: ℎ ≤ 60, 60 < ≤ 240, > 240; Table 4 show the normalized distribution of the car parking events.
Long . We use diferent feature combinations to train the models: (1) Single event-related feature where we train the model using only one event-related feature; (2) All event-related features together where we train the model using all single event-related features at once. We refer to when we use all the event-related features to train the ML model. For both cases, we perform two further combinations: using and not using the spatial and occupancy features to feed the models.
Hyper parameters. We use a grid search to tune the hyperparameters of the algorithms [16]. Specifically: For Figure 2: Distribution of occupancy of the parking lots of XGB, AB and RF, we vary the number of trees in the PKLot (CNR, omittted due to lack of space, has similar figures) range of {50, 100, 150}, while maximum tree depth vary in the ranges of {2,3, or until all leaves are pure}, respectively; For SVM, we use the RBF kernel with varying
Training Approaches. Given the ordinal characteris- in {0.0001, 0.001, 0.01}. For the LR, we have used two tic of our target variable, we have explored three super- diferent class weight parameter {balanced and uniform}, vised approaches to train the predictive models over the while the multi class parameter changing between {auto, training set. For each approach, the best model is selected ovr (for binary classification)}. For the K-means, the taking into account the average results over the 5 folds varies in the ranges of {2, 3,4,5,6}. While for the DBScan of validation. These approaches are: (a) Classification: ranges are {50, 75, 100, 125, 150} and {2,3,4} for the and the training is performed without taking into account minimum sample, respectively. the order of the classes and the selected model is the one Baselines. To be able to evaluate the performance with highest micro-fscore; (b) Regression: the training is of our approach we have used the following baselines: executed to reduce the mean square error (MAE) of the (a) Random: randomly choose a class; (b) Longest Class: predicted values, therefore the model with lowest error always select the longest interval; (c) Shortest Interval: is selected. always choose the shortest interval; (d) Majority Class:
Algorithms. For the Classification and Regression always choose the class with highest frequency in the tasks we used the following algorithms: Random For- training data. For regression, we compare with the (e) est (RF), XGBoosting (XGB), AdaBoosting (AB), Logistic Linear Regression (LN). Naive Bayes and Linear RegresRegression (LR) and Support Vector Machine (SVM). To sion are both simple ML models with high bias. They are compute the spatial features, we have used the K-means used here as baselines given their easy interpretation. and the DBScan clustering algorithms. For all algorithms, ML model training process. For each dataset, we we used the implementation available in the scikit-learn split the car parking events into train and test with 0.8 and library4. 0.2 ratio respectively without shufle the data. To avoid
Features. The following features are extracted and data leakage, we ordered the car parking events using used to feed the ML algorithms. The event-related fea- their timestamps before split. When training the models tures include hour of the day ℎ, time stamp minutes , on the training data, we use a stratified cross-validation day of week , slot id , and weather condition ; with 5 folds. After the training, for each algorithm, the the spatial features include the spatial cluster id ; the best configuration of hyper-parameters is used to retrain occupancy features include the spatial cluster occupancy the model using the whole training data and then assess its performance now using the test set.
Evaluation metrics. To evaluate the experiment re- f1_macro are also recorded in the PUCPR dataset using sults we have used the following measures: micro f1- the classification task. Moreover, we observe that the use score ( 1), macro f1-score ( 1) and mean of the spatial and occupancy features in most of the cases absolute error (MAE). We recall that the F1 score is a (16 out of 24) has improved the performance of the ML weighted average of the precision and recall where best models. value is 1 and worst is 0. The micro f1-score is a metric Altogether, these analysis show an consistent advanwhere we compute an F1 score counting the total true tage in the use of low bias ML models such as the XGB positives, false negatives and false positives. The macro to predict parking event duration due to the implicit ranf1-score is a metric that treats all classes equally, then domness and non-linearity of such events. it does not take label imbalance into account. Indeed, Both classification and regression algorithms produce the macro-average computes the metric independently a similar performance for long term parking; however, for each class and then take the average, hence treating classification is more accurate on the short term forecastall classes equally, whereas the micro-average will ag- ing while regression has an overall lower mean average gregate the contributions of all classes to compute the error in the medium range. The parking prediction modaverage metric. These measures give some clues about ule based on classification could provide a better user the precision and recall of the models on predicting the experience to drivers because accurate identification of true positives. By using the MAE we want to have a short-term parking will force the controller to guarantee more interpretative measure of our regression models a higher energy share to short-term park events. Howsince it computes the average error of the predictions ever, it will reduce the peak shaving capabilities of the values (ˆ) compared to the real values (). For all ex- parking area. On the other hand, a regression model periments, we consider the MAE obtained over the test could facilitate demand response measures because of set as comparison criteria between the models. forecasted parking events shifted towards long parking time. 4.2. RQ: Accuracy of ML in predicting
parking duration
ing the accuracy of our car parking event duration prediction models. At this first study, we analyse the performance of each ML approach ( and ) when predicting parking events duration.
Table 5 report the MAE, micro f1-score and macro f1score of the models. The MAE was used as comparison criteria to select the best models. For each dataset and ML approach pairs, the table indicate the strongest baseline and report the results of the best ML models for two set of features: (a) using only event-related features, represented as {ℎ,,,,,}, having no spatial and occupancy features; and (b) using event-related features with spatial and occupancy features, represented as {,}. The table also report the improvement in percentage achieved by the ML models over the baselines.
We highlight in bold the best MAE result per dataset. Acknowledgements
From the results, we can observe that the ML models overcome the baselines in all the datasets, for all The work is supported by the ERA-NET Smart Energy the training approaches. Specially the ensemble trees System, Sustainable Energy Authority Ireland and Italian models (RF and XGB) show the best results in most of Ministry of Research with project EVCHIP N. ENSGPLUSthe training approaches with XGB showing the best per- REGSYS18_00013. and GA 19/RDD/579-EVCHIP. formance. For all datasets, when using a approach, we observe that the most robust baseline is the Linear Regression (LN), whereas for the approaches References the strongest baseline is the Gaussian [1] U. Duda-Wiertel, A. Szarata, The analysis of Naive Bayes (GNB). In Table 5 the best MAE performance transport-related behaviours of drivers in highly (0.316) is reached by classification with XGB when predicting over the PUCPR dataset. The best f1_micro and
challenges the distribution network’s capacity, and it is becoming a threat to the grid’s reliability. The use of EV batteries as flexible energy storage opens up several research questions on integrating charging vehicles with an Energy Management System. Such a system requires implementing a parking occupancy prediction module, proposed in this paper, based on historical data to forecasting the parking duration for each parking slot. We evaluated various machine learning algorithms across four diferent parking datasets to predict parking behaviour in this context. Future works include the improvement of the current performance results with finer prediction intervals and more classification options. CNR PUCPR GNBℎ GNBℎ
Algorithm RFℎ XGB, SVM SVM XGBℎ XGB, RF RF, RFℎ XGB, XGB XGB,