More than 70 partners from business, science and industry have joined forces for the WindNODE joint project to advance the shift to renewable energies.  The DAI Lab at TU Berlin and an experimental quarter in Berlin’s Prenz­lauer Berg district also participated. The quarter complex includes six buildings with a total of 224 apartments that are heated by a local heating plant. The quarter concept also includes an intelligent energy management system, which enables flexible access to existing renewable energy offerings.

While AI applications can be used to more efficiently deploy locally generated renewable energy and thus conserve resources, the AI applications themselves also consume resources. To ensure effective conservation, AI applications must consume less than they save. AI systems are particularly helpful in producing forecasts for how much energy will be consumed in the quarter at a specific time, and forecasting when and how much energy can be produced via photovoltaic facilities. The diagram on the next page depicts the model of the residential quarter that served as the basis for the simulation study. The forecasts produced by AI ensure that the energy saving system (ESS) and the quarter’s hot water reservoir can be used most efficiently.

Such forecasting systems can exhibit a low level of complexity if, for example, statistical models are used. If they are based on a Deep Learning approach, however, they can be extremely complex, while a traditional Machine Learning approach results in mid-level complexity. The greater the complexity of a forecasting model, the more energy it consumes in the development, training and usage (inference) phases. Computer simulations can provide insight into whether the elevated resource demands of more complex AI systems pay for themselves in the form of a more efficient usage of renewable energies.

The example of the Berlin quarter makes clear that AI systems don’t always live up to expectations. The studies completed as part of the SustAIn project show that the resource consumption of AI models aimed at optimizing the feed-in of renewable energies in the quarter was not particularly high. At the same time, however, the savings achieved by those models were relatively small. The largest savings potential was produced by the infrastructure, especially the energy saving system, which allows for the more flexible use of locally produced energy.

The simulations were performed for three AI systems of differing complexity, all designed to increase the share of renewable energy in the electrical power supply of the residential quarter in Prenzlauer Berg. As the table below shows, the model XGB (XGBBoost – Extreme Gradient Boosting) consumes the least amount of energy. It is based in traditional methods of Machine Learning and consumes a total of just 0.38 kWh during development, training and usage. That makes its consumption even lower than that of the less complex statistical model SARIMA (Seasonal AutoRegressive Integrated Moving Average). As expected, the two Deep Learning models – ANN ­(Artificial Neural Network) and LSTM (Long Short-Term Memory) – consumed the most electricity.

Do AI applications increase the share of renewable energy in the quarter’s electricity supply? As part of the case study, the period from October to December was simulated on the basis of energy consumption data from the quarter. Three different digitalization scenarios were modeled:

An energy management system for the quarter with a low level of digitalization served as the base scenario. Excessive heat from the local heating plant is absorbed by a hot water reservoir. To cover peak consumption periods, a boiler is available. Energy for the quarter is ­produced by a photovoltaic facility. In addition, the effects of the installation of an additional energy saving system on overall energy consumption is analyzed.

Based on Scenario 1, data on the heating behavior of the residents is factored in. This information is used to develop an optimized control strategy for the provision of energy and heat.

In this scenario, the forecasts from the different AI systems are used to optimize energy management.

The results show that the greatest difference isn’t achieved through the deployment of AI, but by using an energy saving system in the low digitalization scenario. By doing so, the share of renewable energy in the quarter’s electricity supply is increased from 45 percent to 58 percent. The use of AI models only produces an additional 4 percentage points, to 62 percent. In months with more hours of sunlight, this share will likely be slightly higher.

In the case of the Berlin quarter, complex AI systems were best at producing forecasts for energy consumption and the possible amounts of photovoltaic power that could be produced. The models used for the simulation were all quite economical, but their benefits were limited. AI applications are able to make the energy sector more sustainable by effectively integrating and distributing renewable energy. But they can only realize their full potential within a modern and intelligent network infrastructure. Furthermore, sufficient sources of locally produced renewable energy must be available and storage technologies are necessary to flexibly incorporate that energy. AI, in other words, is not the sole solution.