A novel attLSTM framework combining the attention mechanism and bidirectional LSTM for demand forecasting

Demand forecasting has become the most crucial part for supporting supply chain decisions. However, accurate forecasting in time series demand forecasting, particularly within supply chain operations, is challenging because of short-term data features, such as limited volume, nonlinear datasets, and near history disturbances. As one of the most promising deep learning models, long short-term memory shows superior performance in extracting implicit patterns from datasets of various areas. Thus, a novel forecasting framework, attLSTM is constructed combining enhanced bidirectional LSTM (LSTM) and self-attention mechanism. The forecasting performance of attLSTM is verified by testing six randomly selected datasets and eight additional datasets with different volumes from a given database. The proposed attLSTM is compared with seasonal autoregressive integrated moving average, support vector machine, random forest, and LSTM through two commonly applied evaluation metrics and a specially designed newsvendor cost model. Extended experiments are conducted on four benchmark datasets from other fields. These analyses demonstrate that attLSTM shows comparable performance in supporting the supply chain demand forecasting and operational decisions. The proposed framework has robust generalization capability in univariate time series demand forecasting.