Investigating Predictability of the TRHR Seasonal Precipitation at Long Lead Times Using a Generalized Regression Model with Regularization

Abstract

Skillful long-lead climate forecast is of great importance in managing large water systems and can be made possible using teleconnections between regional climate and large-scale circulations. Recent innovations in machine learning provide powerful tools in exploring linear/nonlinear associations between climate variables. However, while it is hard to give physical interpretation of the more complex models, the simple models can be vulnerable to over-fitting, especially when dealing with the highly “non-square” climate data. Here, as a compromise of interpretability and complexity, we proposed a regression model by coupling pooling and a generalized regression with regularization. Performance of the model is tested in estimating the Three-Rivers Headwater Region wet-season precipitation using the sea surface temperatures at lead times of 0–24 months. The model shows better predictive skill for certain long lead times when compared with some commonly used regression methods including the Ordinary Least Squares (OLS), Empirical Orthogonal Function (EOF), and Canonical Correlation Analysis (CCA) regressions. The high skill is found to relate to the persistent regional correlation patterns between the predictand precipitation and predictor SSTs as also confirmed by a correlation analysis. Furthermore, flexibility of the model is demonstrated using a multinomial regression model which shows good skill around the long lead time of 22 months. Consistent clusters of SSTs are found to contribute to both models. Two SST indices are defined based on the major clusters of predictors and are found to be significantly correlated with the predictand precipitation at corresponding lead times. In conclusion, the proposed regression model demonstrates great flexibility and advantages in dealing with collinearity while preserving simplicity and interpretability, and shows potential as a cheap preliminary analysis tool to guide further study using more complex models.

Hotspots of Predictability: Identifying Regions of High Precipitation Predictability at Seasonal Timescales From Limited Time Series Observations

“…Toward improving predictive skill, future work could use data‐driven (going beyond predefined indices) machine learning methods that are designed to account for high dimensionality and spatiotemporal dependencies of predictor variables. Such methods have shown considerable potential (e.g., see data‐driven and machine/deep learning methods in DelSole & Banerjee, 2017 ; Liu et al., 2018 ; Giuliani et al., 2019 ; Ham et al., 2019 ; Stevens et al., 2021 ; Gibson et al., 2021 ; Peng et al., 2021 among others), and are suited for investigating seasonal precipitation predictability in nonlinear settings. Further improvements may be possible by combining statistical and dynamical model predictions in a hybrid, post processing setting (Hao et al., 2018 ; Khajehei et al., 2018 ; Khajehei & Moradkhani, 2017 ; Madadgar et al., 2016 ).…”