Improving LSTM forecasting through ensemble learning: a comparative analysis of various models

Abstract

Supply chain management involves managing the entire manufacturing process, from purchasing supplies to delivering the final product. Demand forecasting helps businesses predict future customer demand by analyzing historical data and market patterns. While various papers discuss optimizing models, this research compares several machine learning models, such as ARIMA, SARIMA, and deep learning models like RNN, LSTM, GRU, and BLSTM. It also extends to approaches like ensemble learning with the LSTM model, discussing how ensemble learning can further improve the LSTM model. This paper explores ensemble learning in two ways: a) without model pruning, averaging all generated models, and b) with model pruning, removing underperforming models and averaging top performers. Experiments conducted on a public dataset from the University of Chicago achieved a very low RMSE loss of 9.26 on the LSTM model improved via ensemble learning with model pruning. This ensemble approach with model pruning improved accuracy in predicting future customer demand, and a complete pipeline integrating visualization and a notification system was developed.