Forecasting Ozone Concentration Data: Arima v/S LSTM

Harguna Sood

Abstract:

Forecasting time series data is an important
subject in climate monitoring, weather forecasting and
pollution level estimation. Traditional techniques used are
univariate Moving Average (MA) and Autoregressive
Integrated Moving Average (ARIMA). ARIMA models
have proven their superiority in precision and accuracy of
predicting the next lags of time series. Due to the recent
advancements in computational power of computers we
are able to use data intensive techniques such as deep
learning. The question explored in this paper is that
whether or not the deep learning-based algorithms, like
“Long Short-Term Memory (LSTM)”, is better or not
when compared to the traditional algorithms when used
for time series forecasting of weather data. The study
conducted here shows that ARIMA outperformed
algorithms such as LSTM for short-term weather- related
data prediction. The ARIMA model provided a decent
reduction in error rate when compared to the LSTM
approach. Also, there was noticeable difference in the
overall processing time for both the algorithms with the
ARIMA model finishing first, thereby providing reduction
in the running time required for such type of operations.