Novel wavelet-LSTM approach for time series prediction

Time series prediction often faces challenges due to hidden patterns and noise within the data. This paper presented a novel algorithm that combines wavelet decomposition with long short-term memory (LSTM) networks, providing a distinct method for handling these challenges. The study considered the monthly rainfall data (mm) of India from January 1901 to December 2021. The series under consideration was denoised using maximum overlap discrete wavelet transform (MODWT), followed by long short-term memory (LSTM) modeling on each denoised series. The hyperband search algorithm was employed to identify the optimal hyperparameter combination for each LSTM model, aiding in further fine-tuning the model. The algorithm ended with the implementing of the inverse wavelet transform on the final predictions. In order to evaluate the efficacy of the proposed approach, it was benchmarked against the other established models such as LSTM, recurrent neural network (RNN), and artificial neural network (ANN). The result showed that the proposed model (MODWT-LSTM) significantly outperformed the other benchmark models like LSTM, RNN, and ANN in terms of forecast accuracy. Specifically, in terms of root-mean-square error (RMSE), the proposed algorithm witnessed a gain in prediction accuracy to the tune of 18.5%, 32.8%, and 36.47% than that of LSTM, RNN, and ANN model, respectively. The superiority of the proposed model is further confirmed by use of Diebold–Mariano (DM) test, establishing the hierarchy of model effectiveness as MODWT-LSTM > LSTM > RNN = ANN in terms of predictive performance.