Forecasting lightning is critical due to its significance for public safety, infrastructure resilience, and disaster response. Traditional numerical weather models are computationally intensive and typically designed for long-term forecasting. In contrast, machine learning models enable faster predictions, but require training on large datasets.The use of Machine Learning (ML) methods for lightning forecasting has shown promising results. They can be used for a shorter predictive horizon, in the range of 15 to 60 minutes, but also present some limitations as they require large amounts of historical data for training. Additionally, they can be difficult to interpret, making it challenging to understand the underlying mechanisms driving the predictions. This paper presents a new approach to lightning forecasting that addresses such limitations. The proposed approach identifies key lightning activity patterns through the use of Clustering and Support Vector Machines, bypassing the need for training on historical data. The proposed model was applied to satellite-based lightning data, and its performance was found to be comparable to that of machine learning models, while being computationally more efficient.