Short-term electricity load forecasting based on a novel data preprocessing system and data reconstruction strategy

Accurate forecasting of the electricity load plays a crucial role in the decision-making and operation of the smart grid. The characteristics of load series such as non-stationarity, nonlinearity, and multiple-seasonality make such prediction a challenging task. In this study, a novel combined prediction model based on a feed-forward neural network (FNN) is established to address these challenges. A novel multi-step data preprocessing (DP) system is developed, which comprehensively eliminates the undesirable characteristics in electricity load data by identifying outliers, correcting outliers, and smoothing the data. Random forest is used to select high-impact parameters for electricity load to eliminate irrelevant variables. Moreover, a strategy combining seasonal-trend decomposition based on loess (STL) and Hodrick-Prescott (HP) filtering has been developed. STL is utilized to decompose the load series for exploring the inherent connection between electricity load trends and seasonal variations, while HP is employed to reconstruct the residual series, enabling a more in-depth exploration of the latent trend characteristics. The accuracy of the proposed framework was evaluated using publicly available datasets. The proposed DP-HSTL-FNN model achieved the lowest root mean square error (4.11 KW) and the coefficient of determination (R2) was closest to 1, demonstrating the highest predictive accuracy and reliable performance among all compared models. The results validate the significant advantage of the novel DP system in improving the quality of complex datasets, while confirming the potential of the data reconstruction strategy in enhancing prediction accuracy. Therefore, this novel model can be effectively applied in short-term electricity load forecasting for smart grids.