Spatiotemporal convolutional LSTM with attention mechanism for monthly rainfall prediction

Abstract

Climate has always become part of an essential role in human life on earth. Climate change is a scorching topic because of various prediction efforts to predict the future of the earth's climate on human survival. As one institution that deals with climatology in detail, NOAA publishes a journal on the approach used to generate CHIRPS as a spatiotemporal dataset. The data consist of spatial two dimensions and temporal data from 1981 to 2020. The dataset for the experiment contains 480 pictures, which are 480 months or 40 years of data. K-means clustering is used to collect different characteristics of spatial data based on the Within Cluster Sum Square (WCSS) value using the Elbow method. Generalizing four spatial areas of different characteristic data, Authors have proposed a Convolutional Long Short-Term Memory model to carry out and develop by an attention mechanism. A deep learning architecture designed to learn the dependencies of temporal and spatial data exclusively utilizing convolutional layers and recurrent layers. Two issues with convolutional networks' ability to forecast sequences using historical data are addressed by our suggested architecture: (1) break the temporal order throughout the learning process, and (2) demand that the lengths of the input and output sequences be equal. The method's purpose is to improve the prediction for each previous method and having lower error than other models such as GRU, LSTM, and LSTM-AT. Still, every model reported in this paper has characteristics of how they learn the pattern of the data and visualize the error rate and performance in every spatial point. The Convolutional LSTM-AT model achieved MAE of 67.97, with training time around 2.5 minutes.