How machine learning is monitoring your systems

What is Artificial Intelligence?

Artificial intelligence (A.I.) comes in many forms. The popular idea is that machines will soon ‘think like humans’. A computer could reason, plan, learn and understand our language. There are even those who believe that artificial intelligence will surpass humans in the future. We are not there yet, but in the meantime, there are forms of AI, such as machine learning, that can take over relatively simple tasks from humans.

What is machine learning and how does Solar Monkey use it?

Machine learning is a form of AI that involves building software algorithms that can learn from data in order to perform better. Machine learning is often applied where enough data is available and patterns or judgments need to be made. Machine learning algorithms are used to assess the content of social media and to detect fraudulent transactions by banks. It could also be applied to the output of solar panels, because there too, a large amount of data is required and there is a need to assess it.

At Solar Monkey we have a lot of available data: yield information from your systems. We use AI to predict the performance of systems and notify installers when something is wrong. The judgment we have to make is as follows: are your systems performing as they should? In other words: does the actual yield not deviate too much from the expected yield?

a.i. systeem

How do we currently predict the yield of your systems?

Our research team, led by Jaap Donker, made an analytical model in cooperation with TU Delft which describes the physical properties of solar panels and inverters. Benchmark studies and industry standards are translated into calculations of PV systems in practice. Analytical models are thus based on the physical properties of the components. For example, how a solar panel reacts to radiation and temperature.


These properties are combined in the algorithm that produced the most important prediction: the expected amount of kilowatt-hours on an average day. When we put the predictions of the model (based on actual irradiation) next to the measured output of a system, we can make a statement about the performance of the system. Differences between the model and reality will always exist. When the difference between the observed and the expected performance is too large, two conclusions are possible. There is too great an inaccuracy in the model, or there is something wrong with the system. Of course, we want to avoid the first conclusion.

”We want our installers to be notified only when something is really wrong.”

That is why we are using the enormous amount of data for a machine learning model that can predict more accurately than the analytical models. In this way, we will soon be able to make a more accurate statement about the performance of a solar system.

How will we predict yield with A.I.?

We are now working on a machine learning model. This will be used to answer the same question: “Does my system do what you would expect?” The recipe is as follows:

”We use the yield data coming from your systems to see how systems should perform.”

This form of machine learning is also called supervised machine learning. This means that we already have the right answers (yield in kWh) on days when we also know the conditions (irradiation, system characteristics).

solar measurement graphs


The learning of the model, also called training, is an iterative process in which the model gets better and better at finding the right translation of the circumstances into the answers. Because we have a lot of data, there is a lot of room for the model to train properly. You can compare it to working many hours at school or many training hours on the sports field. With us, these training hours are supervised by our colleague Rohi. This doesn’t take forever, but there will come a point where the model no longer performs better after each iteration. Then it will be time to determine whether you have achieved the desired performance and whether that is enough to apply the model in practice.


We think we have developed a model that can perform better than the analytical model and we want to have it operational before the end of the year. Then we can analyse the performance of your systems with it. But how will we know if we are successful?

”We are successful when we are more precise.”

If our A.I. model can more accurately predict the actual yield, and is less likely to be wrong, you will have a reliable tool to know whether your system is performing as it should. We have an evaluation period of about one year. This way we cover all seasons and can honestly say if the new AI-driven method is more accurate than the old one.

What does the future look like?

The market is developing towards a larger aftermarket. Fault detection is very important in this respect. After all, your customers want to continue to benefit from their investment and will call if the yield falls. That is why our installers are faced with ever-increasing service costs as systems age. We want to minimise these by being able to say with great certainty when something is really wrong. A.I. will be used more and more for this.

Ultimately, we hope to offer our installers more and more advanced services, so they can spend more time on the mission we all share: making households more sustainable and solar energy the standard.