Day-ahead Multiple Households Load Forecasting using Deep Learning and Unsupervised Clustering

The share of Renewable Energy Sources (RES) in modern power systems shows a significant rising trend. Due to RES production variability, as well as the stochastic nature of the consumption side, accurate forecasting models are paramount for grid operation. Load and photovoltaic (PV) generation forecasting models are used in Energy Management Systems (EMS) for optimizing the energy balance between the distribution grid and households (energy communities) with PV and battery systems. Load forecasting difficulty increases with the reduction of the number of observed objects (multiple to individual households), as well as with an increase of the timeseries resolution (daily to 1h or intra-hour). This paper presents a comparison of supervised deep learning models for 24h ahead load forecast at 1h resolution of 12 aggregated households. Raw data is preprocessed, and the resulting dataset contains a total of 286 days with uninterrupted 24h sequences. Hyperparameters of the forecasting models are optimized using Keras Tuner in Python. The obtained results are analyzed and compared before and after using unsupervised clustering as additional input features.