Deep Learning-Based Probabilistic Forecasting of Household Energy Consumption in Smart Grids

Smart grids, enhanced by advancements in electronics and communication technologies, have transformed energy distribution and management, necessitating accurate predictions of energy consumption to optimize grid reliability and efficiency. This paper proposes a probabilistic forecasting approach for one-day and one-week-ahead household energy consumption using deep learning models. Neural Hierarchical Interpolation Time Series (NHITS), Neural Basis Expansion Analysis (NBEATSx), Gated Recurrent Unit (GRU), and Long Short Term Memory (LSTM) models are trained on energy consumption data from households in London, United Kingdom, collected via smart meters. These models are optimized using a multi-quantile loss function to account for forecast uncertainty. The forecasting performance is evaluated using Mean Absolute Error (MAE), Mean Squared Error (MSE), Mean Absolute Percentage Error (MAPE), Average Interval Width (AIW), and Prediction Interval Coverage Probability (PICP) metrics, with prediction intervals set at 80% and 90%. An expanding window cross-validation approach is employed for performance evaluation. The results show that the NBEATSx model outperforms all others in one-day-ahead forecasting, achieving a 1.3% lower MAE, while the NHITS model performs best, with a MAE of 0.986 kWh, for one-week-ahead forecasting. Additionally, NHITS excels in one-day-ahead forecasting with a PICP of 96.10% at the 80% interval and 99.56% at the 90% interval. However, for one-week-ahead predictions, NBEATSx surpasses NHITS, achieving a PICP of 90.48% and 97.40%, respectively.