Efficient deep generative model for short-term household load forecasting using non-intrusive load monitoring

Abstract

Home energy management systems (HEMS) enable key strategies and methods to improve residential efficiency and energy utilization. To make informed decisions, HEMS depend on energy monitoring and forecasting systems. However, short-term load forecasting (STLF) at the household level is challenging due to the high uncertainty in load demand caused by the difficulty of expecting customer behavior. Traditional solutions only use household load data at low temporal resolution and cannot adequately consider the users’ power consumption profile, limiting prediction accuracy. This paper addresses these issues and proposes a two-stage approach for the household STLF that leverages past and future appliance load data. In the first stage, the past and future appliance load data is estimated from the aggregate load measurement through a non-intrusive load monitoring (NILM) approach based on a deep generative model using variational autoencoders (VAE). This granular information is then augmented through a novel household load predictive model. We implement six configurations combining three stage-1 NILM models with two stage-2 predictive models based on VAE and temporal convolutional networks. Given the volatile nature of residential loads, identifying the running appliances and their consumption benefits the household STLF. The proposed approach is designed to leverage short-term appliance load forecasts rather than being limited to historical features, which is a key advantage, especially when multi-state appliances are used. The proposed approach is compared to state-of-the-art approaches on the UK-DALE and REFIT datasets and yields competitive results with improvements of 16% and 19% on average for the MAE and MAPE metrics, respectively.