Adding AI to road lighting
Lighting of modern streets and highways does not only provide better visibility and hence safety to drivers but can also improve the aesthetics of the surrounding areas. However, excessive lighting can also increase environmental pollution due to increased energy consumption and also due to lighting pollution that causes the sky glow effect during the nighttime. It is estimated that more than 80 %of the world’s population and 99 % of Americans and Europeans, live under sky glow. Nocturnal light interrupts sleep and confuses the circadian rhythm—the internal, twenty-four-hour clock that guides day and night activities and affects physiological processes in nearly all living organisms. Studies show that light pollution is also impacting animal behaviors, such as migration patterns, wake-sleep habits, and habitat formation.
Fortunately, the use of dimmable LED streetlights can minimize the above-mentioned impacts while maintaining the safety requirements of modern streets and highways. Extrabit, is in the final stages of developing an adaptive lighting system named IntelliLight, capable of dynamically adjusting the light intensity of LED luminaires based on various factors such as traffic load prediction, weather forecasting and incident detection systems. The project is funded by the Operational Program “Competitiveness, Entrepreneurship & Innovation” (EPAnEK) B (http://www.antagonistikotita.gr/epanek_en/events.asp?cs=2)and is coded as T2EDK-05075.
The IntelliLight system
IntelliLight is an innovative, intelligent software platform, based on machine learning technologies, which determine the brightness of LED luminaires in the open road sections of highways, through the processing of traffic and meteorological data (Figure 1). Using traffic load data generated by inducive loop detectors (or similar vehicle counting systems), the system predicts the average future traffic load, utilizing neural networks and machine learning. The predicted traffic is combined with the expected weather conditions and other events on the highway and through a set of rules the required brightness (or dimming of the LED luminaires) is calculated. Finally, the calculated brightness is given as input to the respective controllers from where the lighting on the highway is regulated.
Figure 1 Using machine learning algorithms, inputs are processed, and an accurate dimming decision is directed to the dimming controllers
From theory to practice
IntelliLight was put on the test as part of a pilot project by Olympia Odos Concession Company SA (https://www.olympiaodos.gr). The pilot was implemented on the Ancient Corinth I/C of the Elefsis – Patras highway. Key objectives that were tested include:
- Ability to retrieve and process traffic volume data originating from the inductive loop detectors of Ancient Corinth’s I/C.
- Ability to accurately predict the traffic volume in intervals of 15 minutes for the following four days.
- Ability to receive traffic related incidents from a video incident detection system installed on the I/C
- Ability to receive meteorological alarms such as heavy rain and low visibility though a newly installed meteorological station
- Ability to classify the predicted traffic volume to predefined dimming levels and send the output to the dimming controllers
- Ability to react to external events and raise the dimming in cases of incidents or bad weather conditions
The results exceeded the original expectations, with a prediction accuracy of more than 96% in a four day period and more than 90% for predictions in a 30 minute window. Moreover, the system was able to consistently recognize external events and respond by raising the dimming levels.
Figure 2 presents the overall architecture of the system that was implemented as part of the pilot phase. Blue rectangles depict the processes that were developed whereas the yellow rectangles represent existing infrastructure. The glue logic was implemented by building custom RESTful APIs that organized and presented in a meaningful way the data to the required components.
Figure 2 The overall system architecture. Communication was performed primarily using RESTful APIs.
The modularity of the system allowed us to cover additional requirements from the highway operator such as dimming in predefined levels instead of linear dimming. Specifically, dimming decisions were made on Annual Average Daily Traffic (AADT) limits. As a result, three plus one dimming levels were used:
- Contractual class which is the maximum possible brightness used on incident detection and bad weather conditions
- Nominal class (Initial dimming level)
- 1st Dimming class (lower traffic volume)
- 2nd Dimming class (minimum traffic volume)
Finally, in the event of low accuracy (less than 85% in a four day period) the machine learning algorithm enters a re-train mode that utilizes recent traffic data to adjusts its decision weights.
The way forward
The findings of the pilot phase greatly helped the development team to understand how the system must function and coexist on a complex infrastructure such as the ones used by highway operators. Additionally, it enabled us to realize the need of modularity and the ability of the system to support a vast array of different systems that influence the final dimming decision. IntelliLight’s ability to adapt and perform in such conditions secured Extrabit a contract with Olympia Odos that will have IntelliLight installed and manage the dimming of all new dimmable LED road lights that are currently installed. The project is scheduled to be completed by early 2023.
For the obligatory screenshots of the IntelliLight user interface, figure 3 depicts the predictions of the ML algorithm with red (30 minutes ahead) and green (1 hour ahead) lines aligned to the actual traffic with a blue line in a period of one day. The data are for the Ancient Corinth I/C. As it is evident, the predictions even for the pilot phase were quite accurate with the 1 hour being much closer to the real-time traffic than the 30 minute prediction.
Figure 3 Traffic prediction compared to actual traffic. The accuracy of the four day prediction is also visible in the right gauge.
Figure 4 presents the final dimming decision for an uneventful night. The output is enabled after a specified time in the afternoon and is disabled in early morning. As it is evident from the diagram, traffic was predicted to rapidly decrease after 23:00 and hence the dimming was lowered to the minimum allowed class (2nd dimming class) at approximately 23:30. On the contrary, traffic was predicted to rapidly increase at 6:00 in the morning and as a result IntelliLight proactively increased the dimming to the Nominal class at 5:30.
Figure 4 Dimming decisions. IntelliLight reactively decreases the dimming level when there is low traffic and proactively increases the dimming level when traffic is predicted to increase.