Comparison of different drought indices and forecasting rainfall by using artificial neural networks

The overall objective of this study was to compare six drought indices and drought forecasting using the ANN in the Punjab Province of Pakistan. Comparison of EDI, SPEI SPI, RD, Z-Score and CZI showed that all these DIs are extremely correlated for similar time periods. The correlation increases at long time period, the indices are highly correlate at 9 and 12 months time period. On the contrary SPI, SPEI, Z-Score, RD and CZI were observed to have small correlation with different time periods. DIs at I-month time period provide wrong calculation of drought features because occasionally a small term extra may have dismissed the lengthy drought and divided a usual drought occasion into two small measured and this may not be suitable. Therefore, while associating drought indices, the drought harshness should be calculated using long time period of the DIs. The time step should be chosen in such a way, so that at least one significant rainfall month is included in the time step. EDI is founded to be better correlated with other DI's for all time steps (except same time step of various DI's). The greatest association of EDI was founded in 9, 12 months' time period of all other DI's. To identify the drought strength for study which was designated that 9 and 12 months time period of DI's is a best choice. Generally, EDI is founded to be best option for calculation of drought features and monitoring of drought form, because of its capability of timely detection of drought onset and realistic quantification of severity of drought events in study area. Three Drought indices the EDI SPEI and the SPI were selected to use for forecasting of drought event in this reign. EDI SPEI and the SPI are significantly correlated for all time period. Moreover 25 different ANN models were tried for every DI's about ten meteorological stations in the Province Punjab of Pakistan with times period from 1-12 months. The best models in all cases are established to contain a DIs value from the representing month of the preceding year. The final, good ANN, for all the EDI, SPEI and SPI were comparatively complex architecture. Three layer networks and maximum of six neurons for a hidden layer appeared to be sufficient for all time's period and all stations. In drought prediction, It was mostly important to guaranteed that precise middle and long-term predictions (with lead times of 3 to 12 months) were produced. The better models formulated in this research the values R2 are 0.86,0.98 and 0.84 and Root mean square error are 0.178157, 0.013828 and 0.163184 Root mean square error is 0.116962,0.195486 and 0158181 for EDI, SPEI and SPI which is suggested of a high prediction precision, mostly in the situation of the EDI. Comparing the EDI, SPEI and SPI predictions were exposed that the EDI is batter then SPEI and SPI. The ANN model has a better performance over all times period. It is capable of exactly forecasting the pattern of dry and wet periods of the EDI. Both models are complex structured with inputs that does not varied with the 12 lead time. This makes both models to be operational purposed. Significance of forecasted model for each index was checked by plotting fitted and actual time series to verify that whether the plotted series is uncorrelated or not. The results associated with tile study shows that, MLPNN algorithm set for each drought indices have potential capability for drought forecasting. Applications of network (MLPNN) architecture on drought indices at selected stations were validated through MAE, the R2 and RMSE.