Short-term rainfall forecasting using machine learning-based approaches of PSO-SVR, LSTM and CNN

Short-term rainfall forecasting plays an important role in hydrologic modeling and water resource management problems such as flood warning and real time control of urban drainage systems. This paper compares the performances of three machine and deep learning-based rainfall forecasting approaches including a hybrid optimized-by-PSO support vector regression (PSO-SVR), long-short term memory (LSTM), and convolutional neural network (CNN). The approaches are used to develop both 5-minute and 15–minute ahead forecast models of rainfall depth based on datasets of Niavaran station, Tehran, Iran. Results of applying the models to all data points indicated that PSO-SVR and LSTM approaches performed almost the same and better than CNN. Subsequently, rainfall events were divided into four classes depending on their severity and duration using K-nearest neighbor method, and a separate forecast model was built for each of the classes. Classification of the events improved the forecast models accuracy where PSO-SVR and LSTM were the best approaches for the 15-minute and 5-minute ahead rainfall forecast models, respectively. Investigating the impact of more predictors on the forecast quality, adding differences of rainfall depths to model predictors improved the accuracy of PSO-SVR approach for the 5-minute ahead forecast model up to 13%. Furthermore, depending on the rainfall event, additional input variables considering rainfall depth fluctuations over shorter time periods than the forecast lead time increased the performances of the PSO-SVR and LSTM approaches between 3–15% and 2–10%, respectively.