Preprocessing of the Initial Data with Discrete Wavelet Transform to Predict Time Series Using LSTM

In this study, we attempted to increase the prediction accuracy of the long short-term memory (LSTM) model by preprocessing of the initial time series data with discrete wavelet transform (DWT). The methane concentration time series was decomposed into one approximating and three detailing components with DWT. These four components along with temperature time series were used to train LSTM models. Two predictive models were trained on the preprocessed data (LSTM&DWT (temperature) and LSTM&DWT (time) models), while two other predictive models were trained on an untransformed data (LSTM (temperature) and LSTM (time) models). The prediction was calculated as the sum of the predictions for each component. In general, the proposed LSTM&DWT (temperature) and LSTM&DWT (time) models outperformed LSTM (temperature) and LSTM (time) models. The LSTM&DWT (temperature) model was the most accurate at all assessed performance indices, which was also marked in the Taylor diagram. The result was than proved by the 48 -hour prediction of CH<inf>4</inf> concentrations built with four LSTM-based models.