Instead of several hundred large-scale power stations, millions of large and small wind power and solar installations will be the heart of our energy supply in the future. At the same time, as e-mobility and electrical heating is booming, the number of consumers is on the rise. Coordinating them all is a complex task, and artificial intelligence can help. Take a look to see where self-learning algorithms are already being used, and what will become possible in the future.
The energy transition is not just changing the way electricity is generated – the system is becoming decentralized and digital. According to model calculations, the millions of solar storage systems, wallboxes and heat pumps could provide a chance to make the grid more stable. A scenario by the thinktank Agora Energiewende (German Energy Transformation Initiative) predicts the capacity of residential storage systems and car batteries capable of feeding energy back into the grid to exceed the capacity of pump storage systems by the end of the 2020ies. But a stable power system will depend on the different systems working together like a swarm of bees. Self-learning algorithms and artificial intelligence are already helping to achieve this.
The amount of electricity flowing into the grid coming from photovoltaic installations is so high that grid operators need to know what to expect – the more accurately, the better. In some regions, operators of solar farms are legally required to control feed-in to match grid capacity. In fact, everyone who markets their renewable electricity directly has to estimate their production as precisely as possible beforehand, because balancing out shortfalls is expensive. Players who optimize their self-consumption using batteries or controllable consumers such as electric vehicles or heat pumps also need information in order to decide when to charge and discharge their storage devices. In all of these applications, artificial intelligence is already being used to forecast and optimize energy flows.
The starting point is the weather forecast. “Machine learning has been playing a crucial role in the irradiation forecast for a long time,” Jan Remund, Head of the Department for Energy and Climate at Meteotest AG, a Swiss weather forecast provider, explains. Meteotest uses satellite images together with a combination of physical models and self-learning algorithms to predict the movement of clouds. “The accuracy for the next few hours is quite high. The longer the timespan, the less accurate the forecasts can be,” says Remund. Meteotest provides special data services that predominantly consist of irradiation and temperature data and can be integrated into the PV systems’ respective monitoring and control software. Companies can use the weather forecasts as a basis for calculating solar irradiation using dedicated software, or can buy such calculations.
The technology is constantly being refined. meteocontrol, a monitoring and forecast specialist company from Augsburg combines the data of various weather forecast providers with their own yield calculations. meteocontrol, Karlsruhe Institute of Technology (KIT) and Deutscher Wetterdienst are working on PermaStrom, a joint research project modelling the effect of aerosols such as ash and fine sand on cloud formation. The goal is to make solar forecasts more accurate. “March 3 and 4, 2021 showed how relevant this can be. On those days, we had a lot of Saharan dust in the air. For those two days alone, the optimized forecast saved Germany around three million euros by avoiding balancing energy costs,” explains Stijn Stevens, meteocontrol’s managing director. Utility companies incur balancing energy costs when generation does not cover consumption within their balancing group, forcing grid operators to close the gap.
Ultimately, grid operators are responsible to ensure that grid feed-in and feed-out exactly match each single second. That is why accurate feed-in forecasts are vital to them. meteocontrol and Fraunhofer Institute for Solar Energy Systems (ISE) are among the providers of such forecasts. Grid operators combine the forecast generation capacity from renewables, consumption forecasts and fossil-based power plant output to predict when and where their power lines and transformers will reach their limits.
TransnetBW from the south west of Germany launched the Stromgedacht app last autumn to make the topic of grid capacity available to the public. When things get tight, the app sends push messages to its users, warning them to postpone or reduce electricity consumption. With around 500,000 downloads, the effect on the grid should still be limited, however. Grid operators will continue to have to buy grid stabilization services such as operating reserve to balance out minor capacity deviations, the targeted shift of power plant output to the area behind the grid congestion (redispatch), or balancing out reactive and active power. Power plant controllers are used to enable large PV systems to support grid stabilization. These controllers are able to receive and carry out commands from the grid operator, or independently control the feed-in of reactive and active power depending on local grid parameters. If the worst case occurs and the grid does get overburdened, power plant controllers will curtail the feed-in power.
The way that most decentralized storage devices and consumers are controlled is based on a different logic. Even though storage devices are often described as taking the burden off the grid, they generally do not as yet communicate with the power grid. What they actually do is to ensure that the share of locally generated electricity that is used on-site is as high as possible.
Battery storage systems are just one element that makes this possible. There are also flexible consumers involved. Heat pumps can charge the heating storage device when the sun is shining, and then the energy can be fed into the heating circuit in the evening. Another possibility is to time charging electric vehicles so that the share of solar power is as high as possible. This requires the system to know the weather forecast and the typical consumption behavior. The way compatibility between energy management systems on the one hand, and consumers and storage systems on the other is achieved varies depending on the manufacturer. Either release signals are triggered or excess electricity is reported.
The systems from Solar-Log show this: They can be combined with various battery storage systems. The controlled shift of charging times based on the weather forecast only works in combination with storage systems by partner company VARTA, however. In line with the conditions for the feed-in tariff under the Renewable Energy Sources Act, this function limits the feed-in peak. Once the battery is fully charged, the solar power is used for a heat pump, for example. A smart heat pump makes this combination particularly efficient. Some devices can integrate the yield forecast for the next few hours sent by the upstream energy management system into their planning. This prevents the heat pump heating the thermal energy storage tank with grid-supplied electricity when enough solar energy can be expected within the next hour.
In the future, all prosumers and flexible consumers will have to grow into an overall network. The electricity market offering flexible rates is a first step into that direction. Some energy management systems are able to include flexible rates into their optimization strategy.
A targeted control process actually allows high-performance consumers such as heat pumps or electric vehicles to stabilize the grid. Companies such as Hive Power already offer software for the intermittent feedback of electricity from car batteries into the grid (Vehicle2Grid). Once again, artificial intelligence is involved: The FLEXO Smart EV Charging software learns when the vehicles are needed and during which times electricity is available to be sold. This should allow owners to earn up to 1,000 euros per year. Depending on the situation, the electricity can be used within the house, or sold on the electricity market.
Artificial intelligence is already being used in many different ways, controlling renewable energy systems and integrating them into the grid. So far, learning algorithms have been concentrating on a limited field. In the future, they will have to learn to share data and respond to each other. The rules must be very clear. Setting these rules is one of the topics the energy industry and governments are currently trying to work out – with no results so far.
