Rule-based BPNN model for real-time IDF prediction of rainfall: Valuable Input for Early Warning Systems

Abstract

Rainfall data sources constitute a vital component of flood early warning systems (EWS), and their inseparability from these systems is evident [1]. However, the information derived from these sources is typically confined to the duration, intensity and peak time for ground-based stations and cloud density and temperature for satellite productions [2]. Therefore, more details into the current rainfall occurrence and predictions regarding its future characteristics can significantly assist real-time flood forecasting systems to perform more accurate and reliable measures [3]. One of the rainfall characteristics that can bring valuable insight into the EWS are return period (RP) or position of rainfall into the intensity-duration-frequency (IDF) curves. This new parameter can offer a more nuanced understanding of rainfall events and significantly enhance the capabilities of early warning systems [4]. In this study, a novel Back Propagation Neural Network model is designed to enhance the accuracy of rainfall predictions in EWS. The model incorporates five rainfall inputs of (1) current Intensity, (2) intensity gradient determined from an intensity library, (3) current duration, (4) current RP determined using rules from the IDF curve library, (5) RP gradient, (6) absolute energy, and (7) anthropic class. The model employs two 5-neuron hidden layers to predict the RP class of current rainfall, i.e. a 5-year or 3-month RP for instance, depending on the desired lead time. To evaluate its accuracy, the model is tested for various time predictions with 15-minute intervals. Subsequently, a real case study of an urban drainage system in the UK is chosen to assess how this additional input enhances previously developed models [3-4]. The results demonstrate that the model excels in predicting the RP for a 2-hour lead time, achieving a performance accuracy exceeding 90%. Moreover, an acceptable accuracy rate of over 75% is achieved for a 4-hour lead time. Additionally, the incorporation of an added parameter into a benchmark EWS results in a 10.8% increase in accuracy for 15-min, escalating to 37.8% for 4-hour lead time. Although the influence of the added parameter may be minimal for near timesteps, its impact becomes significantly more pronounced when dealing with longer lead time predictions, exactly when conventional EWS performance tends to be reduced.