Solar power

May 04, 2020

The major strength of photovoltaic panels is that anyone owning a roof or an open parcel of land can potentially install photovoltaic panels. But their strength is also their main drawback as it is quite hard to assess their geographical deployment and, therefore, their impact on the electrical grid.

Photovoltaic panels mapping

In order to predict the photovoltaic panels production, we tried to localize every photovoltaic installation in the province of Liège. To reach this goal, we built and trained a deep neural network to detect and delimit photovoltaic panels on orthorectified aerial images.

The dataset we used for training was the free of access Distributed Solar Photovoltaic Array Location and Extent Data Set for Remote Sensing Object Identification.

Finally, we detected the panels in the province of Liège using the WalOnMap orthorectified images.

Photovoltaic production forecast

We built two models to estimate the photovoltaic production of the province of Liège.

  1. The first one uses the data computed by our photovoltaic panels mapping (location, area and surface azimuth) and irradiance forecasts to estimate the total output power, using a PyStan probabilistic program.

  2. The second model does not use any panel data and relies on a simple equation (using irradiance forecasts) plugged into a PyStan program to estimate the output power.

The irradiance data used is retrieved from Solcast's API, and we use Elia's PV production measurements to fit on.

Wind power

May 06, 2020

In Belgium, wind power is the 3rd most important electricity production source. Unfortunately, it is also the most variable source, as there are numerous factors that influence the wind speed and direction. As a consequence, forecasting the wind farms production isn't an easy task.

Wind farms production forecast

In this section, we aimed at forecasting the total wind power in Wallonia based on weather predictions.

The model that is showcased here, uses a supervised learning method based on gradient boosting. It is constituted of 3 underlying models :

  1. A 10th percentile gradient boosting;

  2. A least-square gradient boosting;

  3. A 90th percentile gradient boosting.

These three models have been trained from a learning set constituted of the wind speeds and maximum wind gusts observed at the 67 wind farms of Wallonia, together with the total wind power produced in Wallonia. See Dark Sky API and Elia's wind power data.

Note. An alternative based on extra trees has also been developped, yielding more coherent quantiles, at the cost of a heavier model.