In an article published in the journal Nature Communications, researchers from Italy developed a hybrid artificial intelligence (AI) method that can forecast lightning flashes up to 48 hours ahead, using features from a numerical weather prediction (NWP) model. AI-enabled technique surpassed the fully deterministic approach of the NWP model and simple prevalence-based random model in terms of accuracy, reliability, and sharpness.
Study: FlashNet: Advancing Lightning Forecasting with Hybrid AI. Image credit: MarekKijevsky/Shutterstock
Background
Lightning flashes represent significant threats across multiple sectors, including wildfire, aviation, telecommunications, electrical infrastructure, and human life. Accurate forecasts of these extreme events are crucial for decision-makers. However, predicting lightning flashes is challenging due to the complex and nonlinear processes involved in their initiation and propagation.
Current lightning forecasts rely on parameterization schemes within NWP models, which excel in medium to long-range predictions by capturing atmospheric interactions and physical laws. However, these models face limitations such as uncertainties in parameterizing lightning flashes and challenges in resolving fine-scale weather features.
AI-based methods emerged as a powerful tool for handling large datasets and extracting patterns and relationships that might be difficult to identify through traditional algorithms. This technique is a fully data-driven forecasting system, i.e., forecasts made solely in terms of the knowledge of field data. However, the majority of existing AI-based methods for lightning forecasting focus primarily on the nowcasting perspective, predicting lightning occurrences only a few hours in advance.
About the Research
In the present paper, the authors proposed a hybrid AI-enhanced model called FlashNet to forecast lightning flash occurrences up to 48 hours in advance using binary classification. The model utilized features from the high-resolution (HRES) NWP global model integrated into the forecasting system (IFS) of the European Centre for Medium-range Weather Forecasts (ECMWF). These features included two-dimensional fields and vertical profiles of variables related to thermodynamics, cloud microphysics, and convection. These were extracted every 3 hours for each grid point of the HRES model, which has a resolution of about 10 km in the area of interest.
The presented technique was based on a deep neural network with a multi-head structure, where one head processed point-wise data and the other handled non-local data from vertical profiles. The final layers of the two heads were concatenated, followed by two fully connected layers resulting in the output. The output consisted of a probability value between 0 and 1, indicating the likelihood of a lightning flash occurrence at a given grid point and forecast lead time.
FlashNet was trained and tested using lightning flash observations from the Italian lightning observation network (LAMPINET), a network of 15 sensors covering the entire Italian territory with a detection efficiency of 90% and a location accuracy of 500 m. The LAMPINET data was aggregated in time and space to match the HRES grid and converted into a binary variable, indicating the presence or absence of lightning flashes in a 3-hour interval and a 10 km x 10 km area. The resulting dataset was unbalanced, with only 1% of events corresponding to lightning occurrence. To address this issue, several techniques were employed to balance the dataset, including oversampling, under-sampling, and synthetic minority oversampling techniques (SMOTE).
The authors conducted a 3-fold cross-validation, where each of the three years (2019, 2020, and 2021) was sequentially taken as the test set, and the remaining two served for training. They evaluated the performance of the designed model using various metrics, such as precision, recall, area under the curve (AUC), reliability, and sharpness. Additionally, they compared FlashNet’s performance with the HRES model, which provided a deterministic prediction of lightning flash density based on an empirical formula involving convective cloud and precipitation information, convective available potential energy (CAPE), and convective cloud base height.
Research Findings
The outcomes showed that FlashNet significantly outperformed the HRES model in predicting lightning flash occurrence across all metrics, forecast lead times, and cross-validation folds. It achieved a recall peak of approximately 95% within the 0 to 24-hour forecast interval, compared to HRES's 85%. Moreover, it maintained a high recall of around 90% in the 24 to 48-hour forecast interval, surpassing HRES. Furthermore, it exhibited higher precision than HRES, particularly over land, where complex convection mechanisms influenced by orographic effects are prevalent.
The paper also demonstrated that the new model is both reliable and calibrated, indicating that the predicted probabilities align with observed frequencies of lightning occurrence. It emphasized that the developed method consistently produces forecasts within the skill area, surpassing the reference forecast based on prevalence. Additionally, the model exhibited sharpness, enabling predictions of events with probabilities differing from prevalence frequency.
Furthermore, the authors highlighted FlashNet's potential across various sectors and domains impacted by lightning flashes, including wildfire management, aviation safety, telecommunication, electrical infrastructure, and human health. With the capability to provide accurate and reliable forecasts of lightning occurrence up to 48 hours in advance, the AI-based model offers decision-makers valuable insights to plan and implement preventive and adaptive measures to mitigate the risks and impacts of extreme events.
Conclusion
In summary, the novel model is an efficient and effective hybrid AI-enhanced method for forecasting lightning flash occurrence in the medium-range forecast horizon, using features from the ECMWF-HRES NWP model. The new method is robust and generalizable and performs consistently across different years, forecast lead times, and regions. The authors recommended that the hybrid method be extended to longer forecast horizons and other regions of the world, as well as other rare events such as floods and extreme temperatures.
The researchers acknowledged limitations and challenges and suggested that the novel model can be further improved by employing more advanced AI architectures and techniques, such as convolutional neural networks, recurrent neural networks, attention mechanisms, and ensemble methods.
Drought Prediction with Deep Learning