Evaluating Variational Autoencoders for Synthetic Time Series Data Generation in Agricultural and Energy Applications

The use of AI and machine learning provides powerful decision-making tools that help farmers optimize crop productivity and promote sustainability in agriculture. Critically, the utility of AI-enabled in this context is dependent on the availability of historical data about crops and their environment, which can be difficult to obtain. The creation of reliable synthetic data can help address this problem by enabling the use of machine learning in crop management systems where there is shortage of historical data. This study describes an evaluation of Variational Autoencoders (VAEs) for synthetic time series data generation in crop management. The study's results indicate that VAEs perform very well in generating reliable synthetic time series data using a relatively small crop yield dataset as training material. Further evaluation of the VAE synthetic data generation model on the Panama electricity load forecasting dataset corroborates the validity of the approach.