
Open access
Author
Date
2023Type
- Doctoral Thesis
ETH Bibliography
yes
Altmetrics
Abstract
The use of renewable energy has become an essential aspect of global efforts to mitigate climate change and reduce greenhouse gas emissions. However, the integration of renewable energy into the electricity grid poses significant challenges, particularly due to the intermittent nature of many renewable sources. In recent years, artificial intelligence (AI) has emerged as a promising solution to help address these challenges and enhance the transition to renewable energy. AI can optimize energy systems, improve the forecasting of renewable energy resources, and enable better decision-making processes. This thesis explores the various applications of AI in the renewable energy sector and examines their potential to enhance the deployment and integration of renewable energy technologies. Through a detailed analysis of case studies and real-world applications, this thesis aims to provide insights into the benefits and limitations of AI for the renewable energy transition, as well as its broader implications for sustainable energy systems. The main findings of this thesis are threefold:
1. Expert domain knowledge matters regardless of ground-breaking advances in AI:
While AI has shown great promise in optimizing energy systems and improving forecasting without being explicitly programmed, there are many areas where expert domain knowledge is necessary. This is particularly true in domains where data is unavailable or where training surrogate models is desired. In our first contribution, we demonstrate this on an example where AI can be utilized in generating spatio-temporal car parking maps, which are crucial for coordinating vehicle-to-grid applications in power systems that are transitioning towards high shares of electric mobility and renewable energy supply. This use-case is especially challenging, because parking data is currently not directly measured and available to learn a functional relationship from data at scale. Instead, AI methods can infer this data from hand-designed functions that explain the probabilities of driving and choosing a destination based on changes in more easily measured travel time data. In our last contribution, this finding is further highlighted for the training of surrogate weather and climate models, as well as quantum chemistry-based models that emulate density functional theory. However, with the recent breakthroughs in generative modelling and large multi-tasking models, we can expect the demand for collaboration with AI practitioners and training specialized models to shrink instead.
2. "Big data" is good, only if we are also able to identify the most informative data:
The increasing amount of data generated by digitalization is not always helpful. It is essential to be able to identify the best data to use, as otherwise, the vast amount of data available can quickly become overwhelming and even counterproductive. This is especially true for the power sector, where an increasing amount of data is being measured to shed light into our grids that we have so far been able to operate as black boxes. Proper data selection and processing is necessary to ensure that the data used in AI models are reliable and accurate. We make this discovery for a particular case in which a vast collection of candidate data points is provided for predicting the electric load profile of individual buildings across time and space. We illustrate how active deep learning can utilize additional computation to determine which data is most informative and useful, resulting in more accurate predictions with a smaller dataset, compared to traditional passive deep learning methods, which measure and use data points at random and are currently the default method of choice in almost all existing deep learning applications in the power sector and elsewhere. In our last contribution, the importance of active learning is further highlighted by showing that many applications for enhancing renewable energy require an unpractically large number of model parameters to fall into an over-parameterized regime, given the vast amount of data that is commonly measured for training AI models. These are models that are increasingly achieving near-optimal generalization performance on out-of-distribution data across many different application domains.
3. Spatio-temporal processes play a key role for tackling climate change with AI:
A pattern that emerges in many AI applications for tackling climate change is that these have a data component in time and/or space. This can be used to bring a large variety of tasks and data modalities onto a common ground for designing generalist AI models for multi-tasking and transfer learning, which especially benefits solutions in application domains with little to no data available for training specialised models. We make this discovery throughout all our studies, and illustrate it on a diverse collection of machine learning tasks related to enhancing renewable energy in our last contribution. These include the forecasting of electric demand based on satellite imagery, the forecasting of wind power generation of an entire wind farm, the prediction of traffic for electrifying mobility, increasing the resolution of weather and climate models, accelerating the discovery of electro-catalysts for green hydrogen production and combustion, and the analysis of climate and energy policy texts for effective policy designs. By creating these insights, we want to accelerate the development of large, generalist multi-tasking and transfer learning AI models like ChatGPT and Gato in application domains related to tackling climate change. Show more
Permanent link
https://doi.org/10.3929/ethz-b-000641426Publication status
publishedExternal links
Search print copy at ETH Library
Publisher
ETH ZurichSubject
artificial intelligence; machine learning; deep learning; climate change mitigation; global warming; energy transition; HIdden Markov Model; active learning; Spatio-temporal data; spatio-temporal embedding networks; spatio-temporal embedding uncertainty; climate policy; energy policy; policy analysis; electro-catalyst discovery; climate and weather modelling; electric mobility; wind power forecasting; electric load forecasting; power systems; unified data representationOrganisational unit
09451 - Patt, Anthony G. / Patt, Anthony G.
More
Show all metadata
ETH Bibliography
yes
Altmetrics