DSBAL: Distributed Stacked Bidirectional Attention-based LSTM Method for Time Series Forecasting

Abstract

Abstract Accurate time series forecasting is crucial to increase the performance and turnover of every business. However, It’s quite a difficult task due to the non-stationary and high level of uncertainty in the time series data. This paper proposes a new method called the Distributed Stacked Bidirectional Attention Long Short-Term Memory Neural Network (DSBAL) for time series forecasting. The DSBAL method combines the Stacked Bidirectional LSTM (SBiLSTM) and Attention mechanism in distributed computing. The proposed method consists of an SBiLSTM encoder, attention mechanism, and SBiLSTM decoder. SBiLSTM encoder is used to extract the complex features in the daily tomato supply data, in addition, the Attention mechanism is introduced to enhance the performance of SBILSTM by selecting the more appropriate sequence in the data by giving higher weightage to them. SBiLSTM decoder uses the most appropriate sequences from the attention mechanism to predict the daily tomato supply data. The entire process of the proposed method runs in distributed computing to improve efficiency, accuracy, and scalability. Our proposed method allows us to use only appropriate sequences in the data, captures complicated patterns, and addresses computational issues. To prove the efficiency of the proposed methodology, the experiments are conducted with other time series forecasting methods like RNN, LSTM, Stacked LSTM, Bidirectional LSTM, and Attention LSTM using daily tomato supply datasets in terms of SMAPE and RMSE. The results obtained from the experiment demonstrate that our proposed method is more efficient, accurate, and scalable.