New Tool for Better Production Forecasting

The grid operator collects information about each PV installation through the pre-registration and final registration processes. This includes data on the connection point, the DC capacity of the modules and the AC capacity of the inverter. They already have information on cable dimensions and cable length. The grid operator measures the solar energy input and voltage and frequency variations. However, they lack information on the orientation of the PV systems, i.e. azimuth and tilt, says Mr Lindahl:

- In their operational and grid planning, they therefore often assume a worst-case scenario, which means that all systems can produce at their maximum capacity at any given time and that there is no load in the buildings where they are installed. Without information on the orientation of the solar panels, there is no way of knowing how much is actually being produced, leading to inaccurate forecasts.

Johan Lindahl's remote sensing tool, Analemmas, provides the grid operator with information on the azimuth and tilt of solar PV systems. In short, the process works like this:

An orthophoto is divided into many small frames. A machine learning convolutional neural network (CNN) classifier determines whether a given aerial image contains a PV system or not. The classification is manually corrected in a few uncertain cases by visual inspection of the image supported by a cross-check of registered PV systems from the local grid operator (DSO), and in rare cases by visual on-site inspections. To obtain the surface area and exact coordinates of the panels, a polygon covering the panels is required. Until now this has been done manually, but now a U-net model is being developed to automatically segment the panels. The polygons are used to link the PV systems to specific property designations and specific buildings through Lantmäteriet's geodata services. The property designation is the identification key used to link the identified PV systems with the correct PV system in the grid owner's databases. Lantmäteriet's building layer contains 49 detailed building purposes. These allow the PV systems to be categorised into different market segments.

- Once we have collected all the information, we can simulate the production and then we can quantify the smoothing effect that exists between different plants that have different orientations.

Is this a tool that network operators will be able to run themselves?

- No, it is not possible as it requires large amounts of training data to train such a machine learning method, which is time consuming to develop. But we can help them via our already trained remote sensing method to get the data they need to make good calculations in their network planning tools, says Johan Lindahl and continues:

- To take an example, if a grid operator has a substation with high PV penetration and they feel that they are reaching the limit of what is possible to cope with, then we can make calculations and provide them with simulation values that they can use to make calculations that better match real-life production than assuming that all PV plants produce their maximum capacity at the same time. As a result, costly grid investments can be avoided or postponed.

Background:

The European Commission has estimated that a significant amount of renewable electricity could be lost in the future due to insufficient capacity and bottlenecks in Europe's electricity grids. In a 2040 grid expansion scenario developed by the European Commission, up to 310 TWh of renewable energy could be lost due to grid congestion. This is equivalent to current electricity use in countries like France and Germany. The value of this wasted renewable energy is expected to rise from €26 billion in 2030 to a staggering €103 billion in 2040.

Source

Marie Kofod Hansen, Energiforsk