Seq2Seq-LSTM With Attention for Electricity Load Forecasting in Brazil

Abstract

Electricity load forecasting is important to planning the decision-making regarding the use of energy resources, in which the power system must be operated to guarantee the supply of electricity in the future at the lowest possible price. With the rise of the ability of forecasting based on deep learning, these approaches are promising in this context. Considering the attention mechanism promising to capture long-range dependencies, it is highly recommended for sequential data processing, where time series-related tasks stand out. Considering a sequence-to-sequence (Seq2Seq) time series data of the electricity load in Brazil, this paper proposes the use of the long short-term memory (LSTM) with the attention mechanism to perform the time series forecasting. The proposed Seq2Seq-LSTM with attention mechanism outperforms other well-established models. Having a mean absolute error equal to 0.3027 the proposed method is shown to be promising for field applications. The proposed method contributes to time series forecasting by implementing an attention mechanism to Seq2Seq data, therefore, more than one correlated signal can be used to perform the prediction enhancing its capacity when more data is available.

Keywords

• Electricity supply industry
• energy measurement
• forecast uncertainty
• time series analysis