Subseasonal to Seasonal (S2S) Prediction Algorithms Using Hybrid Machine Learning Techniques

Abstract Subseasonal to seasonal (S2S) prediction based on weather forecasts has received significant attention in recent years due to its potential application in various sectors, such as agriculture, energy, and water management. The challenges in S2S prediction stem from the complex and nonlinear interactions between atmospheric and oceanic processes, which can exhibit significant variability in short, medium, and long time scales. Traditional forecasting methods, such as physics-based numerical weather prediction (NWP) models, have limitations in capturing this variability and have shown limited skill in S2S prediction. However, recent advances in machine learning (ML) algorithms offer new opportunities for improving S2S predictions by enabling the identification of complex patterns and relationships in large and diverse datasets. This paper combines predictions from physics-based numerical weather prediction models with historical data to improve the forecast accuracy of temperature and precipitation averaged over forecast durations of 3–4 and 5–6 weeks using hybrid ML techniques. Specifically, a random forest classification model built for each of 23 regions of the world, as determined by a subject matter expert showed an overall 12.3% improvement in ranked probability skill score (RPSS) for temperature predictions and 4.2% for precipitation predictions for a forecast duration of 3–4 weeks; the improvement for forecast duration of 5–6 weeks was modest: 5.7% for temperature and 0.1% for precipitation. Significance Statement Machine learning algorithms have emerged as powerful tools in the weather forecast research, given large and diverse datasets. This study aims to investigate which features are useful for the subseasonal to seasonal prediction with a time scale of 2 weeks to 3 months based on hybrid machine learning techniques. Since the effects of features may vary depending on geographical and environmental conditions, we divide the global forecast area into 23 regions and improve the forecast accuracy in each region. This work has important implications for subseasonal to seasonal prediction and can be improved by using a wider variety of features available in the model for each region.