Lightning potential forecast over Nanjing with denoised sounding-derived indices based on SSA and CS-BP neural network

• We constructed a CS-BP neural network to forecast lightning over Nanjing, China.
• Cuckoo search algorithm was used to optimize the traditional BP network.
• The best sounding-derived indices were determined by a selection algorithm.
• Singular spectrum analysis (SSA) was adopted to filter out noise of the raw indices.
• The forecast performance of the model is proved to be effective and efficient.

The method of using the back propagation neural network improved by cuckoo search algorithm (hereafter CS-BP neural network) to forecast lightning occurrence from sounding-derived indices over Nanjing is presented. The general distribution features of lightning activities over Nanjing area are summarized and analyzed first. The sounding data of 156 thunderstorm days and 164 fair-weather days during the years 2007–2012 are used to calculate the values of sounding-derived indices. The indices are pre-filtered using singular spectrum analysis (hereafter SSA) as preprocessing technique and 4 most pertinent indices (namely CAPE, K, JI and SWEAT) are determined as inputs of CS-BP network by a linear bivariate analysis and selection algorithm. The cases of 2007–2010 are used to train CS-BP network and the cases of 2011–2012 are used as an independent sample to test the forecast performance. Some statistical skill score parameters (namely POD, SAR, CSI, et.al.) indicate that the CS-BP model excels in lightning forecasting and has a better performance compared with the traditional BP neural network and linear multiregression method.