AI for Solar Panel Defects: A Guide for Utilities

Solar panels.

Solar Panel Defects Cause Problems

Defective solar panels are a major problem for utility companies. Defects in panels can ruin solar energy production, and can often require additional maintenance which increases costs. To combat such anomalies, utilities are turning to UAV-based aerial inspection for defect identification. AI/ML can help automate this inspection and increase detection accuracy.

In this article, we will provide a guide to the types of defects in solar photovoltaic panels, methods for automatic inspection, and how the existing approaches can be improved.

Increasing Prevalence of Solar Installations

It is no secret that solar energy production continues to increase worldwide. In 2020, the International Energy Agency (IEA) reported energy generation from renewable sources grew 7%. Utility companies have driven a significant portion of this increase, as utilities often operate the largest solar installations.

Large solar installations are capable of collecting a lot of energy, but defective panels and equipment may limit the potential of an installation. For example, a defective panel with a hotspot can lead to a myriad of defects, some as extreme as fires which can affect an entire section of the installation.

The challenge with inspecting large solar farms is that manual inspection is often too slow, too expensive, or even error-prone. Solar operators need a consistent, effective, and affordable way to inspect their installation.

Automated Solar Panel Inspection

Recently, unmanned aerial vehicles (UAVs or “Drones”) have been used to conduct automatic inspection. A UAV is capable of surveying an entire solar installation in a few hours, recording high-quality video along the way.

This high-quality aerial video captured by a UAV can be processed using AI to automate quality control for solar panels. Aerial inspection is a zero-contact method to examine a large number of solar panels quickly. With the right AI algorithm, utility companies can run inspections quickly and efficiently.

Any operator knows that on-site inspections can be expensive, and that cost can be compounded by the material replacements for defective parts. Even more severe, downtime can limit the Return on Investment (ROI) for solar operators, and therefore it is of the utmost importance for operators to have a low-cost method for frequently inspecting all of their solar panels.

Types of Defects in Solar Panels

Solar panels can have several types of defects, each of which can lower the photovoltaic output (and thus ROI) of a solar installation:

  • Cracks in a Solar Cell
  • Solar Cell Oxidation
  • Disconnection of Electrical Components
  • Bird Droppings

These defects can often manifest in a hotspot, which is an area of a solar panel that does not have a photovoltaic reaction, leading to overheating. This excess heat can crack the solar cell, which is expensive in time and material for solar installation operators.

Fortunately, hotspots can be detected in aerial imagery. UAVs are often equipped with cameras in the visible (VIS) or infrared (IR) spectrum. Images taken by UAVs can be analyzed with Machine Learning to determine whether a defect is present.

Machine Learning is a Superior Approach

Several traditional approaches to defect detection have been tried in the past. These include techniques such as:

  • I-V Curve Tracing
  • Electroluminescence Imaging
  • Manual visual inspection

However, these techniques are not robust for real world deployments, nor scalable or cost-effective. Fortunately, AI and Machine Learning offers a solution: automated inspection by UAVs at low cost, with high accuracy.

Machine Learning (analogous to the technical terms of “Deep Learning” or “Convolutional Neural Network”) has been shown to outperform many conventional image processing techniques. For this reason, utility companies and academic researchers have begun investigating the potential application of machine learning for solar defect classification.

The early results are promising. Researchers from Italy found that UAV-based IR imaging identified hotspots with approximately 85% accuracy, and UAV-based VIS imaging identified defects with 70% accuracy.

Why Launch a Proof-of-Concept

If a utility company is interested in deploying AI inspection to their solar installation, the best way to get started is a proof-of-concept.

A proof-of-concept is a relatively affordable way for an electric utility company or solar farm operator to prove the Return on Investment (ROI) from AI inspection.

It is highly recommended to measure the total cost of the proof-of-concept against the cost savings in time and material. This will give you the expected ROI of a larger deployment. And anytime you are in a meeting with senior management, having ROI figures can bolster your case.

Generally, you should look at a proof-of-concept as the cost of learning. Maybe a proof-of-concept will meet your Internal Rate of Return (IRR) rate, maybe it won’t, but you won’t know that unless you try.

Five Steps to Launch a Proof-of-Concept

Embarking on a proof-of-concept is more than just flying a drone over your solar panel installation. There are several other factors to consider.

1) Consider the Budget

The first step to launching a proof-of-concept is to decide on a budget. At Simerse, we always say that financial planning is strategic planning. Understanding the expected deliverables is crucial for calculating the ROI from a proof-of-concept.

Get an understanding upfront of what you are willing to spend on a proof-of-concept, and stay within that budget for its entire duration. Staying within the financial parameters is one of the best ways to convince your boss that the proof-of-concept was worth the investment.

2) Get a UAV with a Video Camera

To conduct UAV-based AI inspection you need, well, a UAV. There are many providers to choose from, but the most important factor is that the UAV has a camera capable of video capture, ideally at HD (1920x1080p) resolution or higher. Camera stability is also important, and therefore a UAV with a gimbal may be helpful.

3) Get a Pilot or Automated Solution

There are two choices for piloting your UAV: either a human pilot or a computer pilot. Human pilots generally have lower upfront costs, but may be more expensive over the long haul. More recently, drone retailers have unveiled autonomous flying solutions, which lowers the cost of UAV operation.

But the most important factor is which approach will be the safest and most reliable. There are FAA regulations to consider. And the last thing you need is a reckless drone flying around your installation, potentially crashing into valuable equipment. For these reasons, it is smart to consult with a UAV expert or work with a UAV company on your proof-of-concept.

4) Collect Training Data

Training data is the holy grail of machine learning. At Simerse, we have a proprietary method to generate synthetic training data. But most electric utilities will not have access to that special technology in-house, and will need to rely on real-world data collection absent working with Simerse.

Real-world data collection can be expensive, but without synthetic data, it is absolutely necessary. You will need a fly a drone over your solar installation and take pictures and videos. Then, you will need to manually label that data to prepare it for machine learning. Again, if you want to bypass that process, we highly recommend setting up a meeting with our team.

5) Deploy an AI algorithm

The final piece of the puzzle is an AI algorithm combined with a cloud architecture to process the raw data. For utility companies, this can be the most challenging step in deploying a proof-of-concept.

There are several paths you can take to solve this problem:

  1. DIY (Do It Yourself)
  2. Partner with an AI vendor

Frankly, we recommend the partnership option. This is a little bit of self-interest, but also the DIY route is quite complex. Data scientists are expensive, and an electric utility will have to manage data collection, labeling, and more complex tasks. These activities are generally not a core competency of an electric utility.

The partnership route is a lot easier, and likely even more affordable. Simerse for example has expertise in data collection, management, and algorithm deployment. We offer a full suite of product offerings, allowing you to outsource as much or as little of a proof-of-concept as you like.

Simerse can expand on these promising results

Luckily, Simerse has proprietary technology to dramatically increase machine learning accuracy. For utility companies, Simerse can provide a solution that finds solar panel defects using UAVs at a fraction of the cost compared to alternative methods.

One of the challenges of machine learning is training data. Traditionally, an electric utility would need to capture thousands of pictures of both defective and non-defective solar PV cells. However, defective cells are rare. That means it is very expensive and time-consuming for even the largest utility companies to collect sufficient training data.

Simerse can help. We provide both training data and AI algorithms. We also have the expertise to develop proof-of-concepts and large-scale deployments.

Simerse aims to be that partner for utility companies. If you work at a utility company, get in touch to see if Simerse can help.