Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Machine learning-based forecast of CO2 intensity to support electrical load shifting
In this project, we aim at combining decomposition-based approaches with various machine learning techniques to forecast the short-term CO2 intensity time series (for SA). Furthermore, we aim at using this forecast in the implementation of an emission-aware Model Predictive Controller (for MA).
Keywords: Machine learning - Deep learning - Physics-informed Neural Network - Model Predicitve Control - building and energy applications
Much research has been done to forecast electricity prices, and the literature abounds with proposed methods and algorithms to shift both consumption and production to maximize economic outlook. Operating energy storage technologies for stochastic price arbitrage or optimal bidding strategies for the day-ahead market are such examples. On the contrary, the forecast of the CO2 intensity of electricity – meaning, the “dirtiness” of the electricity depending on the production source – is a much less explored yet very interesting topic. The short-term forecast of CO2 intensity provides electricity consumers with the possibility to shift their consumption in time to minimize their resulting emissions by preferably getting the electricity during times of low electricity “dirtiness”. However, the hourly CO2 intensity time series has rapid changing patterns, making the forecasting task challenging. In this framework, deep learning represents a suitable candidate to tackle this challenge.
Much research has been done to forecast electricity prices, and the literature abounds with proposed methods and algorithms to shift both consumption and production to maximize economic outlook. Operating energy storage technologies for stochastic price arbitrage or optimal bidding strategies for the day-ahead market are such examples. On the contrary, the forecast of the CO2 intensity of electricity – meaning, the “dirtiness” of the electricity depending on the production source – is a much less explored yet very interesting topic. The short-term forecast of CO2 intensity provides electricity consumers with the possibility to shift their consumption in time to minimize their resulting emissions by preferably getting the electricity during times of low electricity “dirtiness”. However, the hourly CO2 intensity time series has rapid changing patterns, making the forecasting task challenging. In this framework, deep learning represents a suitable candidate to tackle this challenge.
In this project, we will combine decomposition-based approaches with various machine learning techniques (mainly deep learning) to forecast the short-term CO2 intensity time series with hourly resolution. Datasets will be provided to the student, but we also expect an independent search for additional data sources if needed to improve the model (e.g. addition of physical information through physics-informed neural netowork). Depending on the time available, the forecast is further utilized in developing an emission-aware Model Predictive Control (MPC) that shifts electricity consumption to minimize equivalent emission. Note that the project is suitable both as SA (where the focus will be on the machine learning-based forecasting task), as well as for MA (where the focus will be both on forecasting task + MPC implementation). The project is done in collaboration with EMPA.
Tasks:
For SA:
- Analyze existing decomposition-based approaches present in literature
- Combine decomposition-based approaches with deep learning methods to forecast short-term CO2 intensity time series at hourly resolution
- Explore additional information or physical knowledge to improve forecast (e.g. physics-informed neural network)
In addition, for MA:
- Simulation-based implementation of an emission-aware MPC in building management or electrolyzer control
- Depending on the progress of the work, an experimental validation of the controller may be included
Both:
- Analyze results and report wri
In this project, we will combine decomposition-based approaches with various machine learning techniques (mainly deep learning) to forecast the short-term CO2 intensity time series with hourly resolution. Datasets will be provided to the student, but we also expect an independent search for additional data sources if needed to improve the model (e.g. addition of physical information through physics-informed neural netowork). Depending on the time available, the forecast is further utilized in developing an emission-aware Model Predictive Control (MPC) that shifts electricity consumption to minimize equivalent emission. Note that the project is suitable both as SA (where the focus will be on the machine learning-based forecasting task), as well as for MA (where the focus will be both on forecasting task + MPC implementation). The project is done in collaboration with EMPA.
Tasks:
For SA:
- Analyze existing decomposition-based approaches present in literature
- Combine decomposition-based approaches with deep learning methods to forecast short-term CO2 intensity time series at hourly resolution
- Explore additional information or physical knowledge to improve forecast (e.g. physics-informed neural network)
In addition, for MA:
- Simulation-based implementation of an emission-aware MPC in building management or electrolyzer control
- Depending on the progress of the work, an experimental validation of the controller may be included
Both:
- Analyze results and report wri
Interested students should send an email to the contacts below with an updated CV and transcript of records (MSc and BSc). Students with machine learning (in particular deep learning) experience are particularly encouraged to apply. A good programming maturity in Python is considered a mandatory prerequisite.
Marta Fochesato, PhD Student, Automatic Control Laboratory, mfochesato@ethz.ch
Hanmin Cai, Postdoc, Empa, Urban Energy Systems Lab, hanmin.cai@empa.ch
Supervising professor: Prof. John Lygeros, Automatic Control Laboratory, ETH Zurich
Interested students should send an email to the contacts below with an updated CV and transcript of records (MSc and BSc). Students with machine learning (in particular deep learning) experience are particularly encouraged to apply. A good programming maturity in Python is considered a mandatory prerequisite.
Marta Fochesato, PhD Student, Automatic Control Laboratory, mfochesato@ethz.ch
Hanmin Cai, Postdoc, Empa, Urban Energy Systems Lab, hanmin.cai@empa.ch
Supervising professor: Prof. John Lygeros, Automatic Control Laboratory, ETH Zurich