Abstract:
The developm ent of an optimising model predictive controller for dom estic storage
radiators w as the ultimate goal of this research project. Neural networks are used
to create empirical m odels that are used to predict the likely temperature response
of a room to the charging of a storage radiator. The charging strategy can then be
optimised based on the real-time price of electricity.
Neural network modelling is investigated by looking at the load forecasting
problem. It is shown how accurate neural m odels can be created and
demonstrated exactly how they process the data. Very specific rules are extracted
from the neural network that can model the load to a reasonable accuracy.
An efficient optimisation technique is sought by optimising the charging of a
dom estic hot water tank based on actual consumption data and the pool price of
electricity. Initially genetic algorithms were tried but their w ea k n esses are
demonstrated. A stochastic hill climbing method w as found to be more suitable.
Monetary saving of 40% over the existing E7 tariff w as common.
The modelling and optimisation are brought together in a storage radiator
simulation. There are improvements in cost and electricity consumption over E7
primarily due to the ability to look ahead and avoid overheating.
A prototype neural controller is developed and tested in a real house. The results
are very encouraging.