Please note: The algorithm descriptions in English have been automatically translated. Errors may have been introduced in this process. For the original descriptions, go to the Dutch version of the Algorithm Register.
Web application Impact monitor Bridge opening
- Publication category
- High-Risk AI-system
- Impact assessment
- Field not filled in.
- Status
- Out of use
General information
Theme
Begin date
Contact information
Link to publication website
Link to source registration
Responsible use
Goal and impact
South Holland province plays an important role in traffic flow within the region by operating and maintaining more than 100 bridges. Smart ICT solutions offer new opportunities to finely link up-to-date traffic information to real-time shipping information via sensors and algorithms.
Of these 100 bridges, 55 have sensors that monitor whether the bridge is open or closed. For an ever-increasing number of these bridges, the province faces a challenge to optimally match traffic flow with shipping and increasing traffic density. The province is committed to smarter, safer and more efficient passage to meet these challenges.
In this, the 'Impact Monitor Bridge Openings' project has taken a substantial step in predicting traffic intensities around a number of important bridges in the province of South Holland. The Long Short-Term Memory neural networks used can make predictions up to 21 minutes ahead with a correlation coefficient accuracy of 75-95%. The developed algorithms were deployed within an Azure cloud environment. It has been delivered as a user-friendly web application in which live shipping and traffic information is presented. This information is used by the application to recommend the optimal time for a bridge opening to a bridge operator.
In a next phase, this information can be fed directly back to the traffic network which, for example, allows navigation software to take into account planned bridge openings. In summary, the developed web application uses recent technology in the field of artificial intelligence that allows bridge operators to combine expert knowledge with the latest technology.
The Impact Monitor Bridge Openings is built to support bridge operators in choosing the optimal time for a bridge opening. The application is equipped to predict traffic volumes up to 21 minutes ahead in combination with up-to-date vessel traffic information. The Impact Monitor Bridge Openings is adapted to the behaviour and work process of bridge operators. In the application, the bridge operator sees 3 time slots of 7 minutes at a bridge. The app calculates in which of the three time periods a bridge opening causes the least inconvenience to road traffic. The bridge operator then makes the decision on when and how long to open the bridge.
Considerations
The algorithm does not process any personal data (such as age, gender, income, health, and so on). The information is in no way traceable to a natural person. Moreover, there is always human intervention and the bridge operator still ultimately decides at what time a bridge is best opened. The algorithm itself is only decision-supportive.
Human intervention
The algorithm itself is only decision support. The bridge operator itself makes the final decision. This makes the bridge operator responsible. There is still human interaction in between. The algorithm itself does not decide when the bridge opens and closes.
Risk management
The Long Short-Term Memory neural networks used can make predictions up to 21 minutes ahead with a correlation coefficient accuracy of 75-95%. This is not 100%. The accuracy of the algorithm could be too low, reducing your insight and the ability to make a decision as a bridge operator.
Also, the algorithm could potentially be misinterpreted by the bridge operator. That chance exists, but its impact is very small. Therefore, a pilot is now being done to bring bridge operators into this so that they understand the algorithm properly.
Legal basis
South Holland province has an important role in traffic flow within the region and is responsible for managing, maintaining and operating more than 100 bridges in South Holland. See among others: https://lokaleregelgeving.overheid.nl/CVDR431299/1
Operations
Data
The data used will be read real-time from the National Database on Road Traffic Data (NDW). Data on waterways is not currently included.
- Number of vehicles
- Type of vehicles
- Speed per vehicles
HIG Measurement loop data
There are measurement loops in provincial roads that record traffic flow in number of vehicles, type of vehicle and speed per vehicle. Only a subset of these loops is accessed real-time to the NDW (National Database on Road Traffic Data) - the rest is used for monitoring and evaluation. As not all measurement loops relevant to the project were available in NDW, a raw dataset was requested from the supplier for the three bridges mentioned in the introduction. This contained the most detailed data of all vehicles travelling over a measurement loop, where each line in the dataset ́is a vehicle. For further use of this data, the individual lines were aggregated by minute with the sum of number of vehicles, average speed and average length and numbers by vehicle classes. Here, between ́e ́e and two years of historical data were used to create the forecast model depending on availability per measurement loop.
NDW real-time road traffic intensities:
NDW has up-to-date traffic data of the number of vehicles passing a measurement point per minute.
Central Object Operation Analysis Log Data (COBALD) bridge opening data. COBALD is a system developed for the Province that is linked to the control centre. It contains all historical bridge openings including the moment (start date/time) and duration of the bridge opening. This historical data was used for preparing and analysing the measurement loop data.
Bridge Sense real-time data:
Additional sensors have been installed on most bridges that detect and transmit real-time bridge openings to Blue Wave Connecting and the NDW - who make it available as Open Data. This real-time data is used for current bridge status in the developed web application.
Technical design
LSTM Long short term memory. Deep learning algorithm
Traffic intensity model development Deep learning models are developed in the Keras library (version 2.2.4), an open-source neural network package written in Python. Keras allows researchers to quickly experiment with deep neural networks. Keras had been used in conjunction with TensorFlow (version 1.5.0) as a backend using the GPU variant of the chosen model (section 3) in the experimental phase. A grid-search had been applied parsed from the Scikit-learn machine learning library for Python. All analyses for data cleaning and traffic intensity modelling were performed in Python.
Artificial Neural Networks (ANN) are a set of algorithms inspired by the hugely efficient functioning of the human brain to recognise patterns. An ANN consists of artificial neurons and in recent years can be trained to far depth and complexity efficiently by so-called deep learning algorithms. For time-series data such as in this case study, Recurrent Neural Networks (RNNs) are often a suitable solution because here neurons not only factor in information from previous layers of the network, but also include information from previous time points. A powerful variant of RNNs is the Long Short-Term Memory (LSTM) network which was developed specifically to model long time series. In the work presented here, LSTM was chosen as the final algorithm to predict real-time traffic intensities for three bridges.
First, an LSTM model was trained on historical traffic data where a time series of inputs was used to predict a vector of steps ahead (script 1). In the second, an encoder-decoder LSTM was used (script 2) where the input time series of data is encoded (encode). The decoder uses this encoded data and makes a prediction one step ahead for each element in the output sequence. The difference is small, as in practice both models predict one sequence ahead. The main difference is that an LSTM model encoder decoder model has information about both the previous step predicted and also the accumulated information from the encoder step is taken into account in predicting the output sequence
External provider
Similar algorithm descriptions
- Whereas ordinary traffic lights register by detection loops (sensors in the road) that traffic is passing over, 'smart' traffic lights additionally register via apps what kind and how much traffic is coming. This allows traffic lights to see traffic "arriving" earlier and thus determine how long the traffic light is on green.Last change on 27th of June 2024, at 13:04 (CET) | Publication Standard 1.0
- Publication category
- Impactful algorithms
- Impact assessment
- Field not filled in.
- Status
- In use
- This algorithm falls under Digital Moat. The boating traffic model has been under development since 2022 and will reflect capacity by canal.Last change on 25th of April 2024, at 12:40 (CET) | Publication Standard 1.0
- Publication category
- Other algorithms
- Impact assessment
- Field not filled in.
- Status
- In development
- Intelligent traffic lights can detect various mobilities with sensing loops and push buttons and connect to in-car systems and navigation apps. To see traffic "coming" at an earlier stage and regulate it even smarter.Last change on 8th of November 2024, at 13:26 (CET) | Publication Standard 1.0
- Publication category
- High-Risk AI-system
- Impact assessment
- DPIA
- Status
- In use