Enhancing Regional Weather Forecasts over Complex Terrain: A Hybrid Physics-Machine Learning Approach

Weather model forecasts over regions with complex terrain are often challenged by geographical complexities, both at synoptic and regional scales. Failure to capture changes in atmospheric behaviors due to terrain significantly reduces the accuracy of weather forecasts. Recent advancements in machine learning (ML)--based techniques have demonstrated promising capabilities in reducing modeling bias compared to observations. However, many of the newly developed methods focus on purely data-driven approaches and often rely heavily on the reanalysis weather data. To address this limitation, our work introduces a novel hybrid physics-ML model to improve regional weather forecasts.

The hybrid physics-ML model comprises two distinct steps. Firstly, we introduce a novel temporal 2D-variation model designed to extract dynamic changes over time from reanalysis weather data, such as the High-Resolution Rapid Refresh (HRRR) dataset, allowing the capture of evolving patterns and the intricate nature of long-term dependencies. Subsequently, our approach involves the development of a hybrid physics-ML model that aims to mitigate the phenomenon of a substantial decrease in accuracy as the forecast horizon extends. This model seamlessly incorporates the outputs from both the temporal 2D-variation model and the weather model (e.g. WRF) as inputs, while relying on observations and high-resolution Large Eddy Simulation (LES) model (e.g., CARPARS) outputs as ground truth. Additionally, we present a computationally efficient, gradient-based explanation method tailored to the trained hybrid physics-ML model, allowing us to gain timely and reliable insights into the relative importance of each input variable in driving predictions.

We apply the proposed model for 48-hour weather forecasts over Oak Ridge, Tennessee, a region characterized by complex terrain attributes such as ridge-and-valley zones and nearby mountain ranges, significantly influencing the weather model's predictions. The results reveal that our method enhances weather prediction accuracy when compared to purely data-driven approaches. The proposed explanation method also provides valuable interpretability by quantifying each variable’s contributions, thereby enhancing our understanding of the factors influencing the forecasts. Moreover, by selecting the most relevant variables and eliminating redundant and irrelevant ones, the accuracy, robustness, and generalizability of ML models are notably enhanced. This refinement effectively alleviates potential model bias stemming from data noise and irrelevant information, consequently, leading to improved predictions and enriched insights. This hybrid ML-physics model leverages both physics-based and data-driven methodologies to alleviate the adverse impact of terrain-induced errors, thereby facilitating more accurate and reliable weather forecasts.