Sequential models for time-evolving regression problems with an application to energy demand prediction

In recent years, there has been a dramatic increase in the use of data that are collected over time and hence models with a temporal component, leading to dynamic models, have received increasing attention. The proposed approach uses a general framework which permits many special cases to be considered. Put simply, for each time a parametric observation model is defined with a conditional auto-regressive type model defined relating the parameters at one time to previous parameter values, this is called the evolution equation.

Simulation results will be presented investigating estimator properties considering a temporally changing regression problem with results demonstrating improved estimation. The technique will also be applied to a real dataset examining the changing relationship between ambient temperature and electricity consumption in the UK. The fitted model can then be used to predict future demand based on easily obtained temperature forecast information.