Search

In this paper, we propose a hybrid model that combines wavelet transform and artificial neural networks (ANN) to forecast the time series of hourly average methane (CH4) concentration. In this work, we want to improve the accuracy of predicting the dynamics of changes in the surface concentration of CH4. The model is based on data from environmental monitoring of greenhouse gases on the Belyy Island of the Yamal-Nenets Autonomous Okrug, Russia. The initial data for building the proposed model were obtained in the period July–August 2017. The time series of the CH4 concentration was decomposed using a discrete wavelet transform into five components—one approximating and four detailing components. These five components, together with synchronized series of meteorological parameters (temperature, humidity, and pressure), were used to train ten ANNs—five autoregressive networks with exogenous input (NARX) and five recurrent neural networks (Elman). The forecast was calculated as the sum of the forecasts for each of the five components. The forecast accuracy was assessed using several indices and a Taylor chart. The hybrid approach improved the accuracy of the forecasts by more than 20%. The hybrid model based on the NARX shown the best accuracy. [ABSTRACT FROM AUTHOR]