Probabilistic Forecasting of Energy Demand in a Household

In this study, a probabilistic forecasting model for energy utilisation is presented. A dataset of energy utilisation for one month of the year at one (1) interval is used to train the long-short-term memory (LSTM) network. The 70 % - 30 % split is used to divide the dataset into the training and testing sets. Given that the dataset is stochastic and non-stationary, a need to account for outliers in the performance arose. To ensure the robustness of the model, the Huber MQ loss function is used for the LSTM model. The results obtained show a reasonable performance of the model. This performance was evaluated using the mean absolute percentage error (MAPE) (24.5 %) and mean absolute error (MAE) (0.83).