Developing a seasonal-adjusted machine-learning-based hybrid time‑series model to forecast heatwave warning

Area of Science:

• Environmental Science
• Public Health
• Data Science

Background:

• Heatwaves threaten environmental sustainability and public health, causing water shortages and landscape drying.
• Accurate heatwave forecasting is crucial for early warning systems and disaster preparedness.
• Forecasting heatwave warnings involves analyzing large-scale, high-frequency daily time series data, presenting unique challenges.

Purpose of the Study:

• To develop and evaluate advanced Machine Learning (ML)-based hybrid models for heatwave warning forecasting.
• To address the complexities of high-frequency time series data in heatwave prediction.
• To compare the performance of proposed models against traditional and existing ML approaches.

Main Methods:

• Proposed two algorithms for ML-based hybrid models and seasonal adjusted ML-based hybrid models.
• Integrated Seasonal-Trend decomposition procedure based on LOESS (STL) with time series and ML models.
• Compared developed models with ARIMA, ETS, TBATS, ANN, SVR, Prophet, RFR, and LSTM using 42-year historical daily data.

Main Results:

• The seasonal adjusted ML-based hybrid model (STL-ARIMA-LSTM) demonstrated superior performance.
• STL-ARIMA-LSTM achieved the lowest error metrics: MAE (0.8974), MAPE (2.9232), RMSE (1.1794), MASE (0.3814), and ACF1 (0.0026).
• The model accurately forecasts the number and duration of heatwaves.

Conclusions:

• The proposed seasonal adjusted ML-based hybrid model offers a significant advancement in heatwave forecasting accuracy.
• This improved forecasting enables better planning and implementation of safety measures against heatwaves.
• The study highlights the effectiveness of integrating STL decomposition with ML for complex time series analysis.