The predictive accuracy of a Rainfall–Runoff Neural Network (RRNN) model depends largely on the suitability of its structure. Unfortunately, the procedures for selecting an appropriate structure for the RRNN have not been thoroughly examined. Inclusion of too many input neurons in the RRNN may complicate its structure, and thereby decrease its generalization performance. The objective of this study is to evaluate the potential of a Principal Component Analysis (PCA) method, i.e. by extracting the principal components from lagged input hydrometeorological data, in improving the predictive accuracy of the RRNN. The Darong River watershed located in Guangxi Province of China, with a drainage area of 722 km2, has been selected to demonstrate the PCA method for modeling the hourly Rainfall–Runoff (RR) relationship. Comparative tests on the forecasting accuracy were conducted among the RRNNs configured with both basin-averaged and spatially distributed rainfall information. Experimental results revealed that, when calibrating the RRNNs with spatially distributed rainfall, the RRNNs using the PCA as an input data-preprocessing tool were found to provide a generally better representation of the RR relationship for the Darong River watershed. However, variable results were observed if the neural networks had been calibrated with basin-averaged rainfall.