Years have passed since the appearance of the first home solar panels. This technology, destined to revolutionise the electricity market, did not have the impact we were all expecting. So what happened?
First of all, we have the problem of not being able to generate energy at night and secondly, that of the distance between users and the energy market. It seems like solutions to the first problem are starting to appear through the use of ever more efficient batteries. This can only mean one thing; we need to bring users and the energy market closer together!
To operate in this market, it is essential to know at what price we can sell and how much we want to sell (or on the flip side, buy). This is the idea behind the area known as Energy Forecasting, in which we try to predict values such as energy prices or consumption. These predictions have always been made based on purely statistical techniques (linear regressions) and working with large volumes of energy (predicting the consumption of a whole city, for example).
I'm sure many of you are already thinking in terms like Artificial Intelligence or Machine Learning. Well you're on the right track! If we want to extend the reach of Energy Forecasting to give us the necessary flexibility for users to be able to access the market more directly, we need forecasts that are adapted to said users. It's thanks to artificial intelligence that we can predict the consumption curve of a particular home, as well as the photovoltaic energy it generates if it has the necessary panels.
Let's take a quick look at how we can predict the consumption of one user (consumption breakdown). The first step has to be to define the error, as it is not the same thing to predict the energy consumption as being lower than the actual consumption as to predict is as being higher.
We will therefore define an error that is greater when the consumption prediction is lower than the actual consumption (for those interested in the details, we are now talking about Quantile Regression).
Secondly, we can use consumption data from multiple users so our model learns to interpret the characteristics of a consumption curve and predict this. For this task, the ideal is to use recurrent neural networks, through which we can adapt our predictions to each user. Said neural network will analyse a user's consumption over a week and return an estimate for future consumption (various hours, one day, etc.).
(Quantile regression: We can estimate different curves depending on the guarantee we want to give that there will not be values below said curve).
In order to predict the generation of photovoltaic energy, we will need access to weather forecasts. We can use an initial model that, based on the weather forecast, gives an initial estimate of the generation curve we will have tomorrow.
For this estimate, models such as Random Forest or Gradient Boosting, based on the assembly of decision trees, can create complex rules based on a large number of weather parameters.
Once we have this estimate, we can use a recurrent neural network to correct recurring errors in the initial forecast. Bear in mind that each panel will have its own particular characteristics that are difficult to include in a model, but which will be reflected in the generation curves (for example, at certain times a tree casts a shadow over the panel, etc.). This type of behaviour is what is corrected with a recurrent neural network similar to that used to predict energy consumption.
We have given a basic description of a forecasting system capable of adapting itself to the characteristics of individual users to allow personalised management of energy. All this development has taken place within the Netfficient project (http://netfficient-project.eu/ y https://www.ayesa.com/es/sectores/smart-life/smart-grid/299-netfficient), which we are excited to see is starting to show us how a different energy market could be possible.