Performance LSTM and Prophet for Prediction Time Series with Limited Data: Case Study of Daily Goat Milk Production

The purpose of this study is to evaluate the performance of LSTM and Prophet to predict milk production of dairy goats that have limited data and only one-time dependent variable (univariate). Both models were built on each dataset. There are 3 datasets namely milk production datasets at Prof Farm and Nyx Farm and 1 combined dataset of both referred to as Probolinggo dataset. The methods used in the research include collecting real data, pre-process data, building prediction models and model evaluation. Handling limited data by adding regressors based on the results of correlation analysis before building a prediction model. LSTM and Prophet were built with two scenarios each, namely models built from univariate and multivariate data. The research results from several scenarios show that the LSTM model on the three datasets is better than Prophet with the coefficient of determination (R
²
) test value around 0.7 for Prof Farm and Probolinggo datasets, while the Nyx Farm dataset is 0.3. Nyx Farm has smaller results because the amount of data is less than the other two datasets. The best scenario is generated by the multivariate data as the model experiences significant performance improvement. The three datasets have their own best models due to different data patterns. The selected model is expected to be used as an initial model to be integrated into the initial development of the goat milk business intelligence system model in East Java.