ConvLSTMNet-EMD: leveraging 1D-convolutional encoder-decoder model with self-attention mechanism for improved energy consumption forecasting

Abstract

Providing a constant, economic, and dependable energy source to end users is a difficult and complex task. To overcome these challenges, energy providers rely on precise load forecasting, especially during periods of peak demand. Although existing deep learning models show potential results, they often struggle to predict peak demand values accurately. To improve energy consumption forecasting, this paper presents ConvLSTMNet-EMD, an overall framework designed to enhance the precision of peak demand values in short-term load forecasting. The introduced framework integrates a 1D-Convolutional encoder-decoder architecture with a long Short-Term Memory network (LSTM) to learn local features and global temporal dependencies. Next, the input time series data is decomposed using Empirical Mode Decomposition (EMD) into multiple components by isolating peak-related patterns, so that the sudden demand variations can be effectively captured. Further, a self-attention mechanism (SAM) is used in a 1D encoder-decoder model to emphasize the informative time steps dynamically by minimizing information loss from the relevant decomposed components. As a result, it enhances the peak demand behavior and overall forecasting performance. The experimental results reveal that the ConvLSTMNet-EMD surpasses the current models with the Root Mean Squared Error (RMSE) of 0.069, which is considerably less than the 0.120 of the existing approach.