Individual household load forecasting using bi-directional LSTM network with time-based embedding

Accurate individual household load forecasting is essential for effectively managing energy demand and promoting efficient energy consumption. This study evaluates the performance of various deep learning models for individual household load forecasting, specifically using the Smart Grid Smart City (SGSC) dataset. Feature engineering is conducted, producing two distinct sets of features, and the models' accuracy in predicting individual household loads is assessed using both basic and derived feature sets. The results demonstrate that the T2VBiLSTM model outperforms other models, achieving an average mean absolute percentage error (MAPE) of 74.90% and a root mean square error (RMSE) of 0.433 kW. The incorporation of a wider range of features, including maximum load, minimum load, and load range over time, is emphasized to enhance forecasting accuracy. These features capture long-term load behavior, enabling the models to comprehend complex energy consumption patterns and generate more accurate and reliable predictions. Moreover, limitations in using MAPE as a loss function for load forecasting at the individual customer level are revealed, due to its sensitivity to scale, asymmetry, outliers, and uniform weighting assumption. Alternative loss functions like mean absolute error (MAE) are recommended, as they treat all errors equally and better capture data peaks. While this study focuses on the SGSC dataset, the findings have broader implications for efficiently managing energy demand and promoting energy consumption on a larger scale.