Synthetic Electrical Load Data Generation: A Tabular Generative Adversarial Networks Approach with Statistical and Predictive Modelling Evaluation

Machine learning-based load models have become increasingly popular because of their exceptional performance across a range of applications but the availability of a sufficient amount of high-quality training data is a major need for these models' efficiency. Since gathering such data is time consuming and costly, synthetic electrical load data generation becomes a crucial first step towards efficient energy systems. In this paper, synthetic electricity consumption data for Panama was generated using Tabular Generative Adversarial Networks (TabGAN). The synthetic data generated are evaluated against original data through descriptive statistical measures such as mean, standard deviation, and median. This evaluation ensures that the synthetic data effectively replicates vital statistical attributes of the original data, a crucial aspect for modelling and analysing energy systems. Furthermore, to assess the practical applicability of synthetic data, predictive modelling is also used. Multilayer Perceptron (MLP) Network and Long-Short Term Memory (LSTM) Network models are trained using both the augmented synthetic dataset and the original training set. A thorough comparison of prediction errors between models trained on the augmented and nonaugmented datasets is done to assess how well synthetic data supports predictive modelling.