Applying context in appliance load identification

We investigate the impact of including context features with conventional machine learning models for energy disaggregation. Four types of context features that were broadly categorized as either temporal context or activity based context were individually examined across ten class of household appliance. We demonstrate that all machine learning models using context features in conjunction with traditional power features produced a significant improvement in classification accuracy of up to 38%. This could be attributed to the context features improving the class homogeneity of the feature space. It was also shown that classes were more linearly separable in the combined feature space of context and power features.