A Novel Short Electric Load Forecasting Approach Using a Multivariate Transformer Neural Network and CEEMDAN

Accurate electrical load forecasting is a difficult task due to the nonlinearity and high volatility of time series data. In this research, the “decomposition and ensemble” methodology is used, which studies have shown to be particularly effective in predicting loads. Complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) is used in conjunction with transformer-neural networks (TRNNs), in order to accurately predict the electricity consumption taking into account the temperature. The proposed approach based on CEEMDAN and TRNN is composed of different stages. CEEMDAN first decomposes the complicated raw load data into a few very basic components. Then intrinsic mode functions (IMFs), extracted from the CEEMDAN, are clustered using the K-means technique based on their entropy index and integrated into primary Co-IMFs. Each Co-IMF is individually forecasted based on TRNN. The final forecasting result is obtained by adding all the predicted Co-IMFs. Hourly data for a six-month time series are used to train the proposed approach while its performance is validated by forecasting one week’s data. The results of the proposed approach are more promising compared with those of Vanilla TRNN, the CEEMDAN- long short-term memory (LSTM) neural network, SARIMA, and SARIMAX.