Trafic flow forecasting is fundamental to today's Intelligent Transportation Systems (ITS). It involves the task of learning trafic complex dynamics in order to predict future conditions. This is particularly challenging when it comes to predict the trafifc status for multiple horizons into the future and at the same time for the entire transportation network. In this context deep learning models have recently shown promising results. This models can inherently capture the non-linear space-temporal correlations (ST) in trafic by taking advantage of the huge volume of data available. In this study the authors present a LSTM encoder-decoder for multi-horizon trafic flow predictions. We adopted a direct approach in which the model simultaneously predict trafic conditions for the entire Belgian motorway transport network at each time step. The results clearly show the superiority of this model when compared with other deep learning models. In the workshop, conference attendees will learn how to process and visualize mobility data, obtain optimal features for trafic flow forecasting, build a LSTM encoder-decoder and perform predictions in an online manner.
Trafic forecasting systems are crucial in modern cities. These systems can potentially provide accurate and timely information to public and private organizations by relying on data collected in real-time. These organizations can in turn take action through policy and lead to more sustainable mobility. These aspects have enormous economic, social and environmental implications. Nowadays, thanks to the development of cutting-edge technologies and advances in artificial intelligence, trafic forecasting has reached results never seen before. In particular, the advent of deep learning (DL) models allowed to take advantage of the enormous volume of mobility data to capture the complex non-linear space-temporal relationships governing road trafic. Moreover, they proved to be particularly efective when it comes to predict trafic for both multiple forecast horizons and for multiple streets. In ITS context, the advantages of DL when compared with traditional machine learning models are summarised as follow [8, 9]: • model huge volume of space-temporal trafic data; • perform multi-horizon road trafic predictions on large scale transportation networks; • achieve greater forecasting accuracy; • comply to real-time requirements characterizing online forecasting systems in ITS;
In the recent literature for multi-horizon forecasting, DL methods can be categorised into (i) iterated approaches using autoregressive models or (ii) direct methods based on sequence-tosequence models [7]. The former approaches use one-step-ahead prediction models where each model’s output is recursively fed into future inputs to obtain multi-step predictions. On contrary, direct methods (ii) are trained to explicitly generate forecasts for multiple predefined horizons in a single step. These models generally present better forecasting accuracy than the iterativebased methods. In this paper, we present a tutorial to build and train a Direct LSTM encoder-decoder model. The model performs predictions for trafic data related to the entire Belgian motorway network. The model is tested over two-weeks period in an online fashion. The results show the DL model obtained better predictive accuracy when compared to advanced seasonal persistence model and other deep learning models. The data and models code are available at https://www.kaggle.com/giobbu/. 2
The aim of trafic forecasting is to predict future trafic conditions given a sequence of historical trafic observations. These observations are detected by sensors, such as GPS, radio frequency identification devices, multi-sensors, cameras and Internet technology, that monitor the trafic status of roads in real-time. In our study, at each time step t , the trafic flow is monitored at S street segments of the entire transportation network. Hence, the road network is modelled as multiple parallel time-series and represented by the matrix X[i,k]: x1,1 ... xi,1 ... xt,1 . . . . . . . . .
x1,k . .
. xt,k . . . . . . . . .
x1,S ...
xi,S ...
xt,S where xi,k is the value of trafic flow at time i and street segment k.
The prediction task can be seen as the process of learning a mapping function, f from X previously observed features to Z future streets’ features, f : X → Z . 3
In the study, the goal is to perform multi-horizon forecasting of the trafic flow based on the current and past trafic conditions of the entire transportation network. To achieve so, a direct strategy is implemented for multivariate multi-horizon timeseries forecasts as discussed in Bontempi et al. [2].
At each time step t , we obtain:
(
4.1.1 Sequence-to-Sequence Learning. In deep learning sequenceto-sequence learning (Seq2Seq) is about training DL models to convert sequences from one domain to sequences in another domain. Multi-horizon trafic flow forecasting at transportation network scale can be seen as Seq2Seq learning problem, where the sequence of trafic flow observations from the past is learned to predict the sequence of observations in the future. The use DL models ofers several advantages [4, 9]: • natively support sequence input data (sequence of flow observations); • directly support multiple parallel input sequences for multivariate inputs (multiple street segments); • map input data directly to an output vector ( multi-horizon predictions).
4.1.2 Long Short-Term Memory (LSTM). Long Short-Term Memory (LSTM) network is a special kind of recurrent neural network (RNN), which is capable of learning short and long-term dependencies of the input data [5] LSTM unit is composed of three gates: input, forget and output gate. These gates determine whether or not to let new input in (input gate), delete the information because it is not important (forget gate) or to let it impact the output at the current time step (output gate). LSTMs are now widely used to deal with sequence data for learning the temporal dependency of the space-temporal data.
4.1.3 Encoder-Decoder architecture. The encoder-decoder models present a particular type of architecture particularly efective for Seq2Seq learning problem [9]. Here’s how it works: • Encoder Module: the input sequence of the encoder is fed to the module. This generates an internal state that is passed as the "context" to the decoder. • Decoder Module: it is trained to predict the target sequence, given the input sequence of the decoder and the memory state vectors from the encoder. The encoder’s state allows the decoder to obtain information about what it is supposed to generate. 4.2
At first we frame our prediction problem by diving the timedependent input features into two categories as shown in Figure 1 : (i) the observed inputs (trafic flow of streets) which can only be retrieve at time step t and are unknown beforehand, and the time-based covariate inputs (features such as day-of-the-week at time t ) which can be predetermined.
The Direct LSTM encoder-decoder (D_lstm_ED) Architecture is shown in Figure 2. As the figure shows, the encoder module takes as input all past features (both observed roads’ trafic flows and known temporal features) and encode them into a latent space. This encoded "context" is then fed as internal state to the decoder module together with the known future covariates as inputs. After being trained, the model perform the forecast of the future (multiple) trafic flow values for the whole road network at each time step.
EXPERIMENTAL SETTINGS 5.0.1 OBU Data . As from 2016, all owners of Belgian lorries having a Maximum Authorized Mass in excess of 3.5 tonnes must pay a kilometre charge. Every road user who is not exempt from the toll must then install an On Board Unit (OBU) recording the distance that a lorry travels on Belgian public roads. Because of their value as mobility indicator, the OBU data are made available to Bruxelles Mobilite’, the public administration responsible for equipment, infrastructure and mobility issues in BruxellesCapital Region. Each truck device sends a message approximately every 30 seconds (from 3 a.m. to 2.59 a.m. of the following day). Each OBU record contains an anonymous identifier (ID resetting every day at 3 a.m.), the timestamp, the GPS position (latitude, longitude), the speed (engine) and the direction (compass). Moreover, OBU data includes vehicle data characteristics: weight category (MAM), country code and European emission standards classiifcation of the engine (EURO value). The large volume and the streaming nature of the OBU data required the set up of a big data platform for an eficient collection, storage and analysis [ 3].
5.0.2 Data Processing & Time-based Covariates. Before obtaining the matrix of Figure 1, the OBU data are processed. Firstly, we consider two months period OBU data, from the 1st of January to the 28th of February, 2019, and the major motorways related to Belgium retrieved from OpenStreetMap1 with Module osmnx in Python. Then, we filter the records with respect to their street segments location and resample with a 30 minutes interval. The trucks are considered on the road network if their location falls within the polygons area representing the street segments 2. We took into consideration street segments with an average trafic lfow higher than 10 vehicles/half-hour. The number of street segments (and consequently the number of time series analyzed) amounts to 5187 with 2832 observations (Fig. 5). Finally, starting from Timestamp variable, we create the following covariates: • a Sine and Cosine transformation of the Hours of the Day.
This will ensure that the 0 and 23 hour for example are close to each other, thus accounting the cyclical nature of the variable; • the Days of the Week. Each day of week shows particular pattern of trafic flow; • the Working/Week-end Days. There is a clear diference in trafic flow between the working days and the week-end days.
This features are deterministic and therefore are known in advance for future trafic flow predictions.
5.0.3 Data Preparation for DL models. In order for the DL model to learn, the sequence of observations must be transformed in the form of { nsampl es , nt imest eps , nf eatur es }. Therefore, we transform X to a three dimensional tensor Tt to represent the trafic flow data, t xwxM, where t denotes the total number of samples, w refers to length of sequence observations and M is the total number of features:
Tt = (X1, X2, . . . , Xt )
(
5.0.4 Seasonal Persistence and other DL models. As a baseline to compare our DL model, we consider Seasonal Window Persistence Model (SW). Within a sliding window, observations at the same time and same day in previous three-weeks seasons are collected and the mean of those observations is returned as persisted forecast. As an example, if the data are hourly and the forecasting target is 9 a.m. on Monday, then if the window size is 1 the observation of last Monday at 9 a.m. will be returned as forecast. A window of size 2 means returning the average of the observations of the last two Mondays at the same hour. Moreover, we compare our model with other DL models: two LSTM models that performs predictions respectively with a direct approach (D_lstm) and with an iterated approach (A_lstm) and a LSTM-based seq2seq architecture with iterated approach 1https://www.openstreetmap.org 2OBU Data processing is available at https://www.kaggle.com/giobbu/obu-datapreprocessing (A_lstm_ED) where the decoder’s predictions are reinjected into the decoder’s input.
5.0.5 Forecasting Evaluation and Metrics. In order to fully exploit the properties unique of real-time data, we adopt the so called interleaved-test-then-train evaluation scheme [1]. Each observation is used to test the model before it is used for training, and from this the forecasting accuracy is incrementally updated. Therefore, the model is always being tested on instance it has not seen. To measure the forecasting accuracy of the competing forecasting approaches we use the well-known Root Mean Squared Error, RMSE, and Mean Absolute Error, MAE. For multi-horizon forecasts these metrics are defined as follows:
RMSEi =
MAEi = s Pit=1 PhH=1 (ei,h )2
nH Pit=1 h=1 |ei,h |
PH nH where ei,h is the error of the forecast for period i and forecast horizon h.
5.0.6 Space-Temporal Resolution and Model Seting. We
forecast the trafic flow (vehicles/half-hour) of each street segment
in the Belgian motorway network up to 6 hours ahead, forecast
horizons from h = 1 to h = 12. From the whole data set we hold
the last two weeks (672 observations) for testing our models. The
data must be scaled to values between 0 and 1 before we can use
them to train the DL model. The predictive models are initially
iftted on the training set. Then, the models are incrementally
updated, while subsequently adding new data from the test set
as described in previous section.
(
A_lstm 11.30 7.12 D_lstm_ED 10.66 6.63
A_lstm_ED 11.08 6.93
Table 1: Average values of RMSE and MAE.
5.0.7 Results. The obtained results are summarised in Table 1. The table shows the average value of RMSE and MAE metrics for both models over the testing set. The Direct LSTM encoderdecoder presents better forecasting accuracy for both metrics. Further, since trafic over the road network may be extremely heterogeneous in terms of volatility, we estimated the normalised metrics against the baseline model [6]. For example, N RMSE is the ratio between the RMSE of the predictor and the baseline RMSE. A ratio below 1 denotes that the considered predictor is at least more accurate than the baseline model. In Fig. 3-4, the average value of normalised metrics is computed for all forecast horizons separately. These figures highlight once again the superiority of Direct LSTM encoder-decoder when compared with other DL models especially for long-range forecast horizons. 6
In the workshop, conference attendees will learn how to perform trafic flow predictions at transportation network scale by employing Direct LSTM encoder-decoder model (Fig. 5). A Jupyter Notebook will be provided in Kaggle3 so attendees can copy it and readily run the code and interact.
The demonstration will be organise as follow:
(
DL models.
The authors acknowledge the support of Programme Operationnel FEDER 2014-2020 de la Region de Bruxelles Capitale 5https://github.com/gboeing/osmnx 6https://geopandas.org and icity.brussels11 for MOBIAID Project12. The authors are also grateful to Bruxelles Mobilite for having provided the OBU data necessary for the work.