Generative adversarial networks for simulating household electricity behaviours

The need to assess how much is the home appliances electricity consumption of a population is increased during the recent years. Among the most important reasons, we can find the necessity for energy providers to better anticipate future demand. Also the users can benefit from this information, reducing the impact on the environment by identifying bad energy habits and costly appliances. The lack of labelled household appliances power signatures, useful for Non Intrusive Load Monitoring algorithms, led to a growing effort in literature in this field. Some generative models were created for reproducing the appliances power signatures not bound to users’ habits or time of the day. A simulator based on Markov chains was designed to simulate activities of end-users, but using only one power signature for each household appliance activation. Our work aims at building a Generative Adversarial Neural network-based simulator trained on different data-sets, capable of reproducing synthetic power signatures of different electric appliances and combining them with users’ daily behaviour. Both appliances electric signatures and users’ behaviour were synthetically generated by our model, which is capable of learning 1-dimensional power data as well as daily habits in terms of electric consumption. The samples generated by our model were evaluated in a qualitative way by looking at their similarities with the real electric loads and in a quantitative way by computing Fréchet Inception Distance score. The final outcome of this thesis work is to simulate the electrical consumption of an heterogeneous population in the correct proportions, in terms of number of family members and their occupations over a one year period. Finally, the consumption obtained by our model were compared with two real world data-set at national and European levels, highlighting a good similarity between the real and the synthetic ones, also confirmed by numerical results.