syalabi-seet / adaptive-solar-tree

The purpose of this project is to train an adaptive solar panel robotic system using reinforcement learning combining shadow detection and solar tracking to maximize power outputs

Adaptive Solar Tree

Abstract

Modern solar panel systems have achieved full solar tracking capabilities but lacks the ability to handle power loss due to shadowing. In this work, we present a novel and intuitive method of increasing solar outputs by leveraging reinforcement learning and a 5-DOF robotic arm manipulator design. We utilized in-game cameras to emulate shadow ratios of solar modules derived from current-voltage sensor readings. The reinforcement agent was trained in a simulated environment equipped with geographically-accurate sun positioning and shadow casting using ray-tracing. Our approach showed promising results in simplifying and integrating solar tracking and shadow detection. Publication link

System Architecture

Episodic flow

Environment

Simulation environment was coded in C# using Unity 3D game engine.

Observation space (15 elements)

• Location's latitude and longitude (2 states, float)
• Solar altitude and azimuth (2 states, float)
• Shadow ratio per module (6 states, float)
• Current motor angles in degrees (5 states, float)

Action space

• Target angle of all motors in degrees (5 continuous values)

Rewards

• +1 when all solar panels are within
  • incidence angle limit < 15 degrees
  • shadow ratio limit < 0.1
• -1 when collision occurs
• -0.001 for every time step

Episode length

• 3000 steps