Abstract
Understanding short term future electricity demand is necessary for maintaining a robust and efficient electricity grid. Research on electricity demand forecasting has traditionally ranged in scale from country-wide grid-level down to individual buildings. However, with emerging technologies allowing smart control of appliances, interest in forecasting the electricity demand of individual appliances is increasing. Smart control of appliances can improve energy efficiency and provide beneficial cost-reducing services to an electricity grid.
In New Zealand, electric hot water cylinders which are present in ~90% of households are currently controlled en-mass using ripple-control as part of centralised demand response programmes to reduce demand at peak times. Smart control of individual cylinders based on their forecasted demand could result in a larger reduction in demand with more nuanced control and less chance of households running out of hot water. This thesis compares a number of existing forecasting methods to the problem of forecasting the electricity demand of residential hot water cylinders at the individual household level. The forecasting models selected for analysis in this thesis were chosen based on their suitability according to existing literature, and preliminary exploratory analysis. A range of conventional forecasting models were assessed including autoregressive integrated moving averages (ARIMA), linear regression, and seasonal and trend decomposition with local polynomial regression (STL). In addition, one artificial intelligence model was assessed, namely support vector machines (SVM). A random walk model was included as a naive benchmark model for comparative purposes.
Existing one minute resolution household electricity demand data that was previously collected in the GREEN Grid study was used for forecasting model development. We selected 22 suitable households from this dataset where the demand of electric hot water cylinders was separately metered. Data was averaged over half-hour periods to mimic data currently available from smart meters in New Zealand. The hot water demand data from each household was separated chronologically into 'training' and 'validating' sets. Models were fitted to the training data and tested against the validating data to prevent inaccurate results from overfitting. The models were compared based on (i) forecast accuracy, (ii) computational speed of model fitting, (iii) interpretability, and (iv) ability to replicate underlying physical processes.
A relationship was discovered between the electricity demand of other appliances in the household and future hot water demand, which was incorporated into some models. SVM models were found to be the most accurate, with 16% lower errors than the naive model, however they performed very poorly in other metrics. The most complex conventional model incorporated STL with ARIMA while including other electricity demand as an external regressor. This performed almost as well as the SVM models in accuracy, while also performing reasonably well in other metrics, and based on the analysis in this thesis, would be the recommended model for use in smart control.