Short-term power load forecasting based on EMA decomposition and an improved LSTM-transformer

Abstract

Reliable forecasts of electricity demand over short horizons are required for routine grid scheduling and energy management decisions. Their accuracy, however, is degraded by the coexistence of slow trends, recurring cycles, irregular variations, and measurement noise in load records. This study develops a forecasting architecture that combines exponential moving average (EMA) decomposition, a modified long short-term memory (LSTM) unit, and a Transformer. The EMA front end separates each load sequence into a smoothed baseline and a rapidly varying remainder, so the network does not have to learn both behaviors from an undivided signal. Historical context is encoded with Transformer blocks, whereas the prediction-side sequence is processed by the modified LSTM to retain sensitivity to recent changes. The decoder estimate is subsequently combined with the extrapolated EMA baseline. Tests using public German and United Kingdom electricity-load data show that the complete architecture gives the smallest RMSE and MAE among the evaluated models. The corresponding R² values are 0.9921 for Germany and 0.9826 for the United Kingdom. These results indicate that separating the baseline from the residual and assigning global and local temporal modeling to different modules can improve short-term load prediction.