Accommodating spatial associations in DRNN for space–time analysis

Dynamic recurrent neural networks (DRNN) are neural networks with feedback connections. They are superior to static feedforward neural networks (SFNN) in nonlinear time-series analysis because they can also accommodate temporal associations. However, like SFNN, DRNN presents a black box approach to space–time analysis. This paper seeks to incorporate spatial associations into a DRNN, through its structure and initial weights. It suggests a novel approach to defining the topological structure and initial weights of DRNN based on the spatial associations of spatial units. This is seen as vital for improving the accuracy and efficiency of prediction and forecasting using space–time models. The proposed method is illustrated using three instances of space–time analysis, which are each characterized by different spatial data types (discrete and continuous). Computational accuracy and efficiency are much improved by incorporating spatial associations in DRNN. This reveals that DRNN can be a powerful tool for modeling space–time series with complex spatial and temporal characteristics.