Optimizing Hyperparameters for Time Series Models to Predict Air Quality Using Evolutionary Computation
Batyr Kenzheakhmetov, Guest Student*; Ryan Kaddis, MSCS candidate; Prerak Patel, MSCS candidate; Devson Butani, MSCS candidate; CJ Chung, PhD, Professor
Computer Science
College of Arts and Sciences, Lawrence Technological University; *Astana IT University, Kazakhstan
Optimizing Hyperparameters, Time series, Evolutionary Computation, Deep Learning
submitted by cj424
This study focuses on optimizing hyperparameters for air quality time series models using evolutionary computation algorithms, specifically Evolution Strategies (ES) and the Distributed Evolutionary Algorithms in Python (DEAP) framework. Two deep learning models are employed: Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU). The objective of this model optimization is to identify optimal hyper-parameters for air quality time series models predicting Total Nitrogen Oxides (NOx) concentrations, with the goal of minimizing Root Mean Squared Error (RMSE). The lowest loss of 40.99 was achieved using 115 LSTM units, three DNN layers (with 64, 128, and 27 neurons), 12 lags, and a learning rate of 0.001, optimized via ES (2+2) Evolution Strategies with 1/5th rule. The results highlight the effectiveness of hyper-parameter fine-tuning through evolutionary algorithms in minimizing prediction loss.
CC BY
Back to Search