A Variational Autoencoder-Based Dimensionality Reduction Technique for Generation Forecasting in Cyber-Physical Smart Grids

Modern energy systems often regarded as smart grid (SG) systems are cyber-physical systems (CPS) equipped with advanced metering and smart sensing devices, leading to a high-dimensional data generation in large volumes. To address this challenge, we propose a new variational autoencoder (VAE)- based dimensionality reduction technique for SG data to enable renewable energy generation forecasting with improved accuracy. The proposed method integrates bidirectional long short-term memory (BiLSTM) deep neural networks with variational inference, to generate an encoded representation of the high-dimensional time-series energy data. The encoded data is further utilized as low- dimensional representation of the original data for the application of energy forecasting, which leads to the reduced computational cost and more accurate forecasting results. The efficacy of the proposed VAE-BiLSTM method is evaluated using python programming and tensorflow library on the data traces taken from the Ausgrid solar generation dataset. Moreover, a comparative analysis of the proposed technique is presented with the benchmark autoencoder (AE) and VAE-based methods. Our result analysis illustrates that the proposed VAE-BiLSTM outperforms VAE-RNN, VAE-LSTM, as well as standard AE- based methods using evaluation metrics such as reconstruction error, pinball score, root-mean square error (RMSE), and mean absolute error (MAE).