Forecasting Multivariate time-series data using LSTM Neural Network in Mysore district, Karnataka

Abstract

Advanced and precise forecasting of infectious diseases plays a critical role in planning and providing resources effectively.
Time series forecasting for non-linear issues are accessible using deep learning techniques. The association between
climatic parameters and dengue occurrences was investigated in this work, and a forecasting model was constructed
using a deep learning approach called long short-term memory (LSTM). Univariate and multivariate LSTM time series
forecast models were developed using meteorological and dengue incidence data from January 2006 to December 2019.
For univariate data, the Pycharm/Google Colab platform was implemented, as the deep learning framework Keras,
which is one of the models in the machine learning library based on Tensorflow. The Pandas Python package with builtin
support for time series data was used for multivariate data. The final model was chosen using the mae loss and the
Adam optimizer. Once the model had been fixed, predictions were made using the model. The research showed that
the meteorological factors such as maximum temperature at lag 3, minimum temperature at lag 3, maximum vapour
pressure (lag 0,1 and 2), minimum vapour pressure at lag 1, and vapour pressure daily mean at lag 0,1,4 are all significant
predictors of dengue along with RMSE value of 1.121 . The results indicated that LSTM network has higher prediction
accuracy than any other traditional forecasting methods. Timely management of seasonal diseases such as dengue along
with meteorological parameters can predict epidemics in the future.