BIAS CORRECTION FOR DROUGHT PREDICTION OVER SUMATERA USING ECMWF MODEL

seasonal rainfall from the ECMWF model for drought prediction using the SPI drought index. However, before the ECMWF output is used for prediction, the bias is corrected first. The results of this study indicate that with the bias correction the output can produce better predictions. The strong drought event is generally related to the el-Niño event. The eastern and southern parts of Sumatra are drier and the chance of getting forest fires is greater in August, September, and October which are drier than other months. This condition occurred due to the monsoonal wind. tested the ability of soil moisture to determine the level of drought in South America using the Climate Forecast System (CSFv2) model. The results show that SPI and soil moisture anomaly forecasts are very influential both spatially and temporally.

Frequently forest fire occurred over Sumatra Island require us to provide seasonal forest fire prediction. With this prediction, hopefully, we can manage and anticipate forest fire which will held. This study utilizes the prediction of seasonal rainfall from the ECMWF model for drought prediction using the SPI drought index. However, before the ECMWF output is used for prediction, the bias is corrected first. The results of this study indicate that with the bias correction the output can produce better predictions. The strong drought event is generally related to the el-Niño event. The eastern and southern parts of Sumatra are drier and the chance of getting forest fires is greater in August, September, and October which are drier than other months. This condition occurred due to the monsoonal wind.
- Drought is a significant disaster, but the characteristics of drought are different from other natural disasters because they occur slowly, accumulate slowly, so it is very difficult to identify the beginning and end of this natural disasters (Wilhite, 2010). Drought is starting to form from prolonged water shortages because of an area receives rainfall under normal conditions (Ghulam et al., 2007).
Utilization of climate models for drought prediction has begun to be developed by several researchers. Carrão

Method:-Data and Study Area
The research study area is located on the island of Sumatra geographically located at 6º18 'LU -6º05' LS and 94º08'BT -108º29' BT. The island of Sumatra is a unique land because there is a mountain range that determines the process of rainfall formation around the range. On the island of Sumatra there are more or less 1,111 rain statios that their data can be used for rainfall research (see Figure 1).

Bias Correction
Bias correction of ECMWF daily rainfall data make used the method utilized byPiani et al. (2010) can only correct data distribution, it cannot correct the rainfall events. The first step of the method to do the correction is to identify the type of distribution of probability and probability of rainfall from the station and ECMWF, both of which follow the gamma probability distribution with the Probability Density Function (PDF) calculated using the following equation: 439 Where: x = average daily rainfall (mm) a = parameter of gamma distribution b = gamma distribution scale parameter Γ = gamma function The second step is to create a relationship between the transfer function of the gamma Cumulative DistributionFunction (CDF) between the station rainfall data and the ECMWF rainfall data.
The third step is to determine the transfer function y = f (x) which can be either a linear or polynomial regression equation to correct ECMWF rainfall data. Simulations conducted by Jadmiko et al. (2017) shows that the regression equation that produces corrected rainfall closest to station rainfall is the 3rd order polynomial regression equation with the intercept value returned at point (0.0) (forcing intercept to zero) with the following form of the equation: In this study, the authors will also consider location, lag time in determining the transfer function. After the type of transfer function is determined, the transfer function is used as a correction factor to correct ECMWF data by entering ECMWF data into the transfer function equation so that a corrected ECMWF is obtained.

Results and Discussion:-Validation Model
Comparison between ECMWF rainfall (blue) and rainfall observation (red) is shown in Figure 2. The ECMWF model simulation data is longer than the rainfall observation data, where ECMWF data is available starting in 1981 while rainfall observation data appear to be available starting in 1998. The ECMWF model input can be extrapolated from observations of weather parameters in other places so that even if there is no rainfall observation in a place, the ECMWF model can show the rainfall at that location. The phase ofthe observational datais almost the same as the phase of the rainfall of the ECMWF. Increase in observed rainfall can be well predicted by ECMWF. The basic difference only lies in the amplitude of the rainfall. In some instances, it appears that the ECMWF rainfall is higher than the observed rainfall. The spatial distribution of the relationship between ECMWF rainfall and observational data is shown in Figure 3. The correlation index in the southernarea of Sumatra looks higher than in the northern region. Correlation reaches 0.3 -0.8 in southern while in the northern it is only 0.2 -0.4. This difference occurs caused by differences in the factors that formsthe rainfall in those two regions.

441
Monthly correlation lag of ECMWF data with rain gauge data is shown in Figure 4. The results show that lag 0.5 has the best correlation with other lags. Utilization of the latest ECMWF data is needed to obtain the best correlation results. Figure 5 is the cumulative distribution function of the rain gauge, ECMWF, and corrected ECMWF data. It appears that, after the ECMWF data was corrected, the cdf graph moved closer to the observational cdf graph. This bias correction significantly increases performance of the ECMWF.  Figure 6 shows that the variation in rainfall corrected ECMWF model looks better than the one not corrected. The maximum peak rainfall is slightly lower than the peak rainfall before corrected, overestimate events can be reduced. From Figure 7 it can be seen that the island of Sumatra has a tendency to strong level of meteorological droughts. This can be seen from the fluctuation of the SPI index which only experiences two positive peak values (≥2) which represent wet to very wetconditions. Meanwhile, the SPI index reached several negative peak values (<-2) which indicate very dry conditions, even between 2015 -2016 the SPI index value reached negative values in several months. In that yearsa very strong el-Niñophenomenon occurred which resulted in a rain deficit in most parts of Indonesia, especially in Sumatra. In that year, there was also a massive forest fire disaster, especially in the island of Sumatra.

Bias Correction
443 Fig 8:-Monthly spatial SPI. Figure 8 is a spatial map of the drought that occurred, the more negative (purple), the stronger the drought. From the picture it appears that the strongest drought occurred in August. This is because along with the emergence of the dry season that usually occurs.

Conclusions:-
Bias correction can change the performance of ECMWF monthly rainfall prediction for the better. SPI drought index results show that in the Sumatra region it was dry when El-Nino occurred.In general Sumatra is dry in August, September, and October which is attributed with the impact of the change of monsoon winds.