Hybrid model combining LSTM with discrete wavelet transformation to predict surface methane concentration in the Arctic Island Belyy

Abstract

In this paper, we propose a hybrid model that combines wavelet transformation of the raw data and an artificial neural network with long short-term memory (LSTM). This model allows researchers to increase the accuracy of time series forecasting.The model is based on data from environmental monitoring of greenhouse gases on the Belyy Island of the Yamal-Nenets Autonomous Okrug, Russia.The raw data for building the proposed model were obtained in July–August 2016. The CH4 concentration time series was decomposed using a discrete wavelet transform into four components - one approximating and three detailing. These components, along with a timed temperature series, were used to train six ANNs—two exogenous input autoregressive networks (NAR) and four LSTM networks. The forecast was calculated as the sum of forecasts for each of the components. Forecast accuracy was assessed using several indices (mean absolute error (MAE), root mean square error (RMSE), mean square relative error (RMSRE), Willmott's agreement index (IA1, IA2)) and a Taylor diagram. The hybrid model based on LSTM showed the best accuracy. The hybrid model based on LSTM showed the best accuracy. MAE, RMSE, and RMSRE errors decreased by more than 70%. In terms of IA1 and IA2, the models improved by 11% and 30% (when comparing the best hybrid models LSTM and NAR. For the hybrid LSTM model errors RMSE, RMSRE, and MAE are more than 20% less than for the base LSTM model for training in which data without wavelet transform were used. Willmott agreement indices (IA1, IA2) rose from 32% to 59%, depending on the models being compared. So, the increase in accuracy after applying the described approach was up to 79% for models based on LSTM (more precisely, from 20% to 79%, depending on the indicator). Applying wavelet transform data to train the NAR-based model reduced MAE, RMSE, and RMSRE errors by 27%, 30%, and 31%, respectively. The Willmott agreement indices (IA1, IA2) rose from 41% to 45% respectively. The Taylor diagram also shows the advantage of the proposed approach.