Post-processing Sub-seasonal to Seasonal Climate Forecasts Under Climate Change

© 2021 Yawen ShaoFor managing the impacts of climate variability and change, climate outlooks on sub-seasonal and seasonal timescales are attracting more interest from climate-sensitive communities, such as water resource management, agriculture, and energy. With a profound knowledge of the sources of climate predictability, modelling techniques are rapidly developed for forecasting future climate conditions. Recent advancements are dynamical global climate models (GCMs), which typically integrate atmosphere, land surface, ocean, and sea ice components to comprehensively simulate earth climate system and output a wide array of climate forecasts. However, GCMs usually suffer from long-standing modelling issues, such as systematic errors and the failure of reproducing the observed trends in seasonal climate forecasts. Statistical post-processing techniques are frequently employed to improve forecast performance. Many commonly used methods are found to be effective at removing biases, maximising forecast skill, and improving forecast reliability in terms of ensemble spread, but they are seldom designed to resolve the trend disparity issue in the post-processed climate forecasts. This issue should not be neglected as global and regional land surface temperature and precipitation have shown discernible temporal trends over recent decades. To address this gap, the overarching objective of this thesis is to develop and demonstrate the merit of a new, trend-aware forecast post-processing method that eliminates the trend disparity between climate forecasts and observations while making the forecasts bias-free, skillful, and reliable. The first part of this research aims to develop a new statistical post-processing method to embed the observed trend into seasonal temperature forecasts. I extend the capability of a calibration method, the Bayesian joint probability (BJP) modelling approach, by introducing a new trend component into the algorithm. The new model (named BJP-t) is applied to calibrate January mean forecasts of daily maximum temperatures from the SEAS5 seasonal forecasting system, operated by the European Centre for Medium-Range Weather Forecasts (ECMWF) in three test stations in Australia. In these cases, the BJP-t calibrated forecasts are shown to accurately reproduce the observed trends, and are more skillful, more reliable, and sharper than raw and BJP calibrated forecasts. In the BJP-t model, the trend is entirely inferred from the training data. In practice, given limited available periods of retrospective forecasts for model training, these inferred trends are subject to large sampling errors, and may not reflect true underlying trends in the observations. Accordingly, the second part of my thesis further develops the BJP-t model to account for trend uncertainty. The extended trend-aware forecast post-processing method is applied to SEAS5 seasonal mean minimum and maximum temperature forecasts, and the evaluations are upscaled to the Australian continent. After trend-aware post-processing that deals with trend uncertainty, forecast trends are more robustly inferred than the BJP-t model. Compared to the BJP calibrated forecasts, embedding trends lead to greater forecast accuracy in regions where observed trends are significant or where observed trend direction is wrongly represented in the BJP calibrated forecasts. The third part of my thesis aims to extend the trend-aware method for post-processing seasonal forecasts of precipitation, which is also a key variable for agriculture and water resource management. Several modifications are made in the model algorithm and evaluation tools to cater for the special features of precipitation amounts, including zero occurrences, highly positive skewness, as well as higher variations and larger uncertainty than temperature variables. I apply this improved trend-aware method to calibrate SEAS5 seasonal precipitation forecasts over Australia. Evaluations show that the trend-aware calibrated forecasts properly reproduce observed trends over the hindcast period of 36 years. In some regions with significant observed trends, skill improvements against the BJP calibrated forecasts are evident by embedding trends into the forecasts. Overall, in most regions, the trend-aware calibrated forecasts outperform raw forecasts with respect to bias, skill, and reliability. Operational sub-seasonal climate forecasts are produced by GCMs configured not dissimilar to seasonal forecast models, but little attention has been paid to explore the ability of the sub-seasonal forecasting systems to capture the observed trends. The fourth part of my thesis firstly aims to investigate this question. Preliminary results show that the same trend disparity issue exists in the 20-year weekly averaged retrospective temperature forecasts from the ECMWF extended-range forecasting system, particularly beyond the first week. Subsequently, I adapt the trend-aware method to calibrate and correct the trend in sub-seasonal forecasts. I modify the method to embed a 30-year climate trend into the 20-year calibrated forecasts. The embedded trends are therefore robustly representative of long-term climate changes and overcome the problem that trends inferred from a shorter period may be subject to large sampling variability. Calibration is applied to the ECMWF sub-seasonal minimum and maximum temperature forecasts for Australia with forecast horizons of up to 4 weeks. Results reveal that raw week-1 forecasts exhibit trends consistent with the 20-year observations in many regions while raw week-4 forecasts do not show the trends of the 20-year observations during the hindcast period. After trend-aware post-processing, trends in calibrated week-1 forecasts are roughly aligned with the 20-year observations across Australia, because when raw forecasts are inherently skillful, the trend-aware calibration transfers raw forecast skill and embeds the 20-year apparent observed trends into the calibrated forecasts. For comparison, calibrated week-4 forecasts exhibit the trends of the 30-year observations, because when raw forecasts do not have much skill, the trend-aware calibration reverts the forecasts to the 30-year observed climatology with trends. In general, the trend-aware calibrated forecasts are more reliable than raw forecasts, while being as skillful as or more skillful than raw forecasts for all lead times. The new trend-aware forecast post-processing method shows robustness for resolving the trend disparity issue for GCM sub-seasonal and seasonal climate forecasts. Wider applications of this method have the potential to deliver quality forecasts and build user confidence in deploying the forecasts for decision-making in a changing climate. Further research will adapt the trend-aware method for other hydrometeorological variables