Hyper Parameter Optimization of Stack LSTM Based Regression for PM 2.5 Data in Bangkok

Abstract

Particulate Matter pollution with the magnitude of 2.5 microns is raising concerns from the most thickly populated cities around the world. There are various studies conducted on predictive analytics over the years. Deep learning has emerged as a new technology which is transforming the face of solving classification and regression problems. Various Long Short Term Memory based architectures are proposed in the past to predict time series data. Randomization of activation and optimization functions was done and the best performing combination is selected. Stack LSTM with this selected configuration on the PM2.5 data is found to be better than the existing LSTM based architectures. Inclusion of Adamax optimizer and fine tuning the activation functions in the LSTM layers gave better performance. The performance metrics reported in this paper are evident enough that the proposed architecture with optimized hyperparameters obtained by randomization of layers is found to perform with lesser error rates and training loss.