R-CAE-Informer Based Short-Term Load Forecasting by Enhancing Feature in Smart Grids

As renewable energy usage increases, power systems become more intricate and demand fluctuations intensify. Accurate short-term load forecasting (STLF) is vital for balancing energy supply and demand. Traditional models often struggle with long input sequences, risking critical feature loss due to inadequate capture of long-period characteristics. To address these challenges, we introduce the R-VAE-Informer, a novel forecasting approach that combines Deep Residual Networks (DRN) and Convolutional Autoencoders (CAE). This model leverages the ProbSparse self-attention mechanism and attention distillation techniques to manage the quadratic complexity of long load sequences more efficiently. It also incorporates a Residual Convolutional Autoencoder (R-CAE) to produce two-dimensional representations of load data, thereby enhancing feature representation and mitigating potential feature loss. Furthermore, the TempoLocode submodule integrates time and positional data, effectively capturing long-term dependencies and periodic variations in time series. Tests on two public datasets demonstrate that our model achieves a Mean Absolute Percentage Error (MAPE) as low as 0.97%, confirming its strong generalization capabilities. Overall, this model markedly improves the precision and sophistication of STLF.