Advancing Precipitation Prediction Using a Composite of Models and Data
- Author(s): Pan, Baoxiang
- Advisor(s): Hsu, Kuolin
- et al.
Advances in numerical weather forecasts have brought forward considerable societal benefits and raised expectations for higher resolution, more accurate, and longer predictions. Despite the consistent progresses achieved, the prediction of precipitation remains a less satisfyingly tackled task, with skills falling far behind those of other atmospheric variables. This dissertation serves as an inspection of prediction capacity and an exploration of predictability for the precipitation process, with a particular focus on the region of West Coast United States.
The sources of predictability, accuracy requirements, and optimal model configurations are distinct regarding the considered forecasting scales and ranges. To identify the successes and deficiencies in predictions and benchmark further advances, a seamless assessment of precipitation prediction skill for short range up to subseasonal scale range is conducted. The evaluation is based on the Subseasonal-to-Seasonal Prediction Project retrospective forecast database. The prediction skill–lead time relationship is evaluated, using multiple models, and measured by both deterministic and probabilistic skill scores. Results show advantageous deterministic skills for the evaluated models at Week-1. The best-performing models achieved r ≈ 0.6 for Week-2 predictions.
The potential sources of predictability at extended range from some of the key climate variations are investigated based on a composite of statistical evidences and numerical predictions. Results show that periods of heavy precipitation associated with ENSO are more pre- dictable at the extended range period. The excessive precipitation and improved extended- range prediction skill during ENSO periods are attributed to the meridional shift of baroclinic systems as modulated by ENSO. Through examining precipitation anomalies conditioned on the MJO, I verified that active MJO events systematically modulate the area’s precipita- tion distribution. Most of the evaluated models are still struggling to represent the MJO or its associated teleconnections, especially at phases 3–4. However, some models do exhibit enhanced extended-range prediction skills under active MJO conditions.
The advantageous precipitation prediction skill for short to medium range originates from a steady accumulation of scientific achievements in (1) inferring atmospheric initial states, (2) resolving atmospheric fluid dynamics, and (3) approximating unresolved atmospheric processes. Evaluation results suggest that we have not fully realized the potentials of these advances in fostering a corresponding improvement in precipitation prediction. Here, the old art of forecasting by reading weather chart and advances in deep learning for image recognition are combined to shed light on the precipitation prediction task from a top-down, data-driven viewpoint. A deep convolutional neural network (CNN) model is trained to learn precipitation-related dynamical features from the surrounding dynamical and moisture fields by optimizing a hierarchical set of spatial convolution kernels. The model applies an “end-to- end” learning strategy to automatically search, synthesize, and extract salient spatial features from the resolved high-dimensional atmospheric field for accurate precipitation estimation at daily scale. Experiments for different regions across the contiguous United States show that, provided with enough data, precipitation estimates from the CNN model outperform the reanalysis precipitation products, as well as the statistical downscaling products using linear regression, nearest neighbor, random forest, or fully-connected deep neural network.
The idea of “end-to-end” learning for inferring unresolved precipitation process based on resolved atmospheric field is further explored for hourly scale quantitative precipitation fore- cast. Hourly precipitation observations from various sources are collected, quality controlled, and concatenated to compose a unique long-term (1980/1/1- 2018/12/31) high temporal resolution precipitation observation dataset. A general framework for statistically modeling of spatiotemporal data and making use of inconsistently available observations is developed.
Hourly precipitation predictions using the deep neural network model give r ≈ 0.8 at 2◦ ×2.5◦ spatial scale, while the baseline numerical model achieved r ≈ 0.5. The best performance at hourly, gauge-point scale reaches the order of r ≈ 0.6 for some gauges. However, there is high skill variance in estimating precipitation at such a stringent spatiotemporal resolution. To further test the proposed model in practical forecasts, dynamical retrospective forecast experiments for two atmospheric river land-falling events are carried out using the Weather Research and Forecasting (WRF) model. The WRF dynamical simulations are used to force the trained neural network model for alternative precipitation process predictions. Simulation results verified the consistency and robustness of the proposed approach. It should be noted that the methods here are not intended to replace precipitation-related parameterization schemes using a “black box” model, rather, the target is to set a benchmark for precipitation prediction from a data-driven perspective, and offer directions for improving precipitation related parameterizations.
Overall, this work conducted a systematical evaluation of precipitation prediction skills across a spectrum of critical scales and ranges. Sources of predictability at subseasonal scale are explored based on a composite of statistical analysis and numerical prediction. The potential of deep learning for seeking evidences in improving precipitation prediction is explored by combining high quality observation data with numerical dynamical predictions.