Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Data-driven hourly energy demand prediction
The goal of this master thesis is to develop a physics informed data-driven building energy demand prediction model. Physics-inputs to the model are generated from a fast and simplified building simulation tool called Hive.
Keywords: Building Energy; Sustainability; Simulation; Physics-informed; Machine Learning; Time series; Data-driven
HIVE (Fig.1) is an open-source Rhino Grasshopper plug-in for early-stage architectural design developed by the A/S group. It calculates monthly building energy demand based on the Swiss norm SIA 380. However, for several contemporary challenges regarding variable renewable energy integration, especially solar energy via Photovoltaics, an at least hourly temporal resolution is required. Amongst others, these include solar energy management and electric battery controls optimization considering arbitrage, curtailment, demand response and peak shaving. It reflects current developments in a renewable grid, where dynamic pricing mechanisms are already rolled out in various Swiss cantons, and will be mandatory in Germany by 2025.
HIVE (Fig.1) is an open-source Rhino Grasshopper plug-in for early-stage architectural design developed by the A/S group. It calculates monthly building energy demand based on the Swiss norm SIA 380. However, for several contemporary challenges regarding variable renewable energy integration, especially solar energy via Photovoltaics, an at least hourly temporal resolution is required. Amongst others, these include solar energy management and electric battery controls optimization considering arbitrage, curtailment, demand response and peak shaving. It reflects current developments in a renewable grid, where dynamic pricing mechanisms are already rolled out in various Swiss cantons, and will be mandatory in Germany by 2025.
Based on previous work conducted at the A/S group (Fig.2), the objective is to develop a data-driven (machine learning, ML) model for hourly heating and cooling energy demand prediction. Inputs to the ML model to be explored include building characteristics and envelope properties from Hive (schedules, geometry, material properties, etc.), meteorological info (i.e., weather file), as well as fast simulation outputs from Hive (monthly energy demand and hourly solar irradiation). The ML model should be directly implemented into Grasshopper.
**Methodology/Tools**
1. Machine Learning with Python, using existing libraries for time series forecasting, e.g. skforecast or Prophet. Primarily LSTM-based architectures, but open to any suitable for time series forecasting.
2. Grasshopper implementation using GhPython components and Hops, or ideally a fully deployable Grasshopper plug-in based on C#.Net.
3. Training data: SIA demand simulations and solar simulations using Hive as inputs. Hourly building energy simulations using EnergyPlus as ground truth, executed on the Euler cluster (script existing).
**Expected Outcomes**
A data-driven model for hourly heating and cooling energy demand prediction, given inputs from Hive (monthly energy demand, building features, hourly solar irradiation). Implemented model in Rhino Grasshopper, e.g. using Hops.
Based on previous work conducted at the A/S group (Fig.2), the objective is to develop a data-driven (machine learning, ML) model for hourly heating and cooling energy demand prediction. Inputs to the ML model to be explored include building characteristics and envelope properties from Hive (schedules, geometry, material properties, etc.), meteorological info (i.e., weather file), as well as fast simulation outputs from Hive (monthly energy demand and hourly solar irradiation). The ML model should be directly implemented into Grasshopper.
**Methodology/Tools** 1. Machine Learning with Python, using existing libraries for time series forecasting, e.g. skforecast or Prophet. Primarily LSTM-based architectures, but open to any suitable for time series forecasting. 2. Grasshopper implementation using GhPython components and Hops, or ideally a fully deployable Grasshopper plug-in based on C#.Net. 3. Training data: SIA demand simulations and solar simulations using Hive as inputs. Hourly building energy simulations using EnergyPlus as ground truth, executed on the Euler cluster (script existing).
**Expected Outcomes** A data-driven model for hourly heating and cooling energy demand prediction, given inputs from Hive (monthly energy demand, building features, hourly solar irradiation). Implemented model in Rhino Grasshopper, e.g. using Hops.
Please send your short CV and recent transcript of records to Dr. Christoph Waibel (waibel@arch.ethz.ch).
Please send your short CV and recent transcript of records to Dr. Christoph Waibel (waibel@arch.ethz.ch).
