Recurrent LSTM network versus feedforward classical approach to the time series prediction

Abstract

The paper presents a comparative analysis of the recurrent deep Long Short-Term Memory (LSTM) networks and the classical feedforward approach in forecasting the time series. LSTM neural networks, unlike feedforward supervised networks (Multilayer Perceptron - MLP, Radial Basis Function - RBF, Support Vector Machine - SVM), belong to the family of deep recurrent structures using a modified neuron model called the LSTM cell. The recurrence built into the cell structure allows for long memorization of the short patterns, ideally suited for forecasting processes in which there are close connections between successive time records. The work focuses on forecasting the 24-hour ahead atmospheric pollution levels for such pollutants as PM10 dust, SO2, NO2, and ozone O3. The advantage of the recurrent approach over classical feedforward applications will be shown.