Application of a Hybrid CEEMD-LSTM Model Based on the Standardized Precipitation Index for Drought Forecasting: The Case of the Xinjiang Uygur Autonomous Region, China

Abstract

Accurate forecasting of droughts can effectively reduce the risk of drought. We propose a hybrid model based on complementary ensemble empirical mode decomposition (CEEMD) and long short-term memory (LSTM) to improve drought prediction accuracy. Taking the Xinjiang Uygur Autonomous Region as an example, the prediction accuracy of the LSTM and CEEMD-LSTM models for the standardized precipitation index (SPI) on multiple timescales was compared and analyzed. Multiple evaluation metrics were used in the comparison of the models, such as the Nash–Sutcliffe efficiency (NSE). The results show that (1) with increasing timescale, the prediction accuracy of the LSTM and CEEMD-LSTM models gradually improves, and both reach their highest accuracy at the 24-month timescale; (2) the CEEMD can effectively stabilize the time-series, and the prediction accuracy of the hybrid model is higher than that of the single model at each timescale; and (3) the NSE values for the hybrid CEEMD-LSTM model at SPI24 were 0.895, 0.930, 0.908, and 0.852 for Fuhai, Kuerle, Yutian, and Hami station, respectively. This indicates the applicability of the hybrid model in the forecasting of drought.

1. Introduction

Drought is one of the most serious meteorological disasters, causing increasing damage to agricultural production, economic operations, and modern life [1,2]. Furthermore, the frequency and intensity of widespread droughts is expected to increase. Easy-to-compute drought indicators are often used to monitor and assess the extent, duration, and impact of drought [3,4,5,6]. Due to the wide range of applications and the different understandings of drought in different disciplines, various drought indicators have been proposed and applied in drought prediction.

Common drought indices include the standardized precipitation index (SPI) [7,8,9], standardized precipitation evapotranspiration index (SPEI) [10,11,12], Palmer drought severity index (PDSI) [13,14], and reconnaissance drought index (RDI) [15,16]. Among these, the SPI can be used for drought analysis at various timescales. Furthermore, it has high accuracy in drought classification and can be calculated using only precipitation data [17,18]. As the SPI is widely used in drought research and can be calculated using only precipitation data, it was chosen for use in this study to illustrate the applicability of models in drought prediction.

Precipitation data and the SPI time-series have non-stationary and non-linear characteristics, but traditional data-driven models, such as the autoregressive and moving average (ARMA) model, cannot predict non-linear data very well. Therefore, artificial neural networks (ANNs) have been used in the study of drought prediction [19], and good prediction results have been obtained. With the development of machine learning, the long short-term memory (LSTM) network has been applied in the time-series prediction of drought and its related fields [11,20,21]. LSTM has been proposed to solve the long-term dependency problem and thus has advantages in handling sequences with long intervals and delays [22]. However, a single model is prone to a local optimum in the prediction of time-series, and the prediction effect is not satisfactory. Therefore, many scholars have introduced the decomposition of signals in the prediction of time-series [23,24]. Signal decomposition can extract the local features of the sequence and make the sequence stable. Scholars in related disciplines have decomposed time-series using empirical mode decomposition (EMD) [22], ensemble empirical mode decomposition (EEMD) [25], and complementary ensemble empirical mode decomposition (CEEMD) [26]. These decomposition methods reduce the complexity of the original time-series and improve the predictability of the data. The original series is decomposed into a set of more stable components and a trend term. Among the above three decomposition methods, CEEMD solves the modal mixing problem of EMD and the residual white noise problem of EEMD. Johny et al. [27] used multivariate EMD-LSTM for predicting monthly rainfall in India, and the prediction results showed that the model is superior to ANN hybrids in the prediction of mean and extreme rainfall.

As mentioned above, with the non-linear and non-stationary characteristics of precipitation data and the advantages of LSTM in long time-series prediction, a hybrid model is proposed to improve prediction accuracy in drought prediction. Taking the Xinjiang Uygur Autonomous Region as an example, the main objectives of the present study are as follows: (1) to characterize drought conditions using the SPI at 1-, 3-, 6-, 9-, 12-, and 24-month timescales; (2) to develop the LSTM and propose a hybrid model by combining it with the advantages of CEEMD in dealing with non-linear, non-stationary time-series; and (3) to evaluate the efficiency of the LSTM and CEEMD-LSTM models based on the Kruskal–Wallis test, statistical criteria, and the Kriging interpolation method.

2. Materials and Methods

2.1. Study Area and Data

The Xinjiang Uygur Autonomous Region is the largest provincial-level administrative region in China in terms of area. It ranges from 73°40′ E to 96°18′ E and from 34°25′ N to 48°10′ N, with an average altitude of about 1000 m (Figure 1). The region is far from the sea and deep inland, with low precipitation levels and a dry climate. The average annual precipitation in Xinjiang is about 150 mm, but there is a large difference between the annual precipitation in the northern and southern regions, with only 20–100 mm in the southern region and 100–500 mm in the northern region. Furthermore, the temperature in the south of Xinjiang is higher than that in the north, with an annual average temperature of 10–13 °C in the south and one below 10 °C in the north. Xinjiang was used as the study area due to its large area. Furthermore, the main purpose of this study was to investigate the applicability of CEEMD-LSTM for drought prediction. Therefore, four meteorological stations located in the northern, southern, eastern, and central parts of the region were selected as examples to display the SPI calculation values and the predicted values of the two models. The original data were obtained from the National Meteorological Data Center (http://data.cma.cn/ accessed on 13 March 2020).

2.2. Methods

2.2.1. Standardized Precipitation Index

The SPI was developed by McKee et al. [28] and has been recommended by the World Meteorological Organization [29]. It is calculated based on precipitation data only. SPI values on 1-, 3-, 6-, 9-, 12-, and 24-month timescales can be used to describe meteorological drought, agricultural drought, and hydrological drought conditions in a region [30,31,32]. The calculation process of the SPI is as follows:

$$\mathrm{SPI}=B \left(t-\frac{c_0+c_{1}t+c_2{t}^2}{1+d_{1}t+d_2{t}^2+d_3{t}^3}\right),$$

(1)

where B is the positive and negative coefficient of probability density; for $B=-1$, $t=\sqrt{ln\frac{1}{G{\left(x\right)}^2}}$, and for $B=1$, $t=\sqrt{ln\frac{1}{{\left(1-G\left(x\right)\right)}^2}}$, where $G\left(x\right)$ is a cumulative probability [33]. The values of the constants are $c_0$ = 2.515517, $c_1$ = 0.802853, $c_2$ = 0.010328, $d_1$ = 1.432788, $d_2$ = 0.189269, and $d_3$ = 0.001308. The drought classification based on SPI values is shown in

Table 1. Drought classification based on standardized precipitation index.

SPI Value	Classification
−0.5+	No drought
−0.5 to −0.99	Mild drought
−1.0 to −1.49	Moderate drought
−1.5 to −1.99	Severe drought
−2.0 and less	Extreme drought

.

2.2.2. Complementary Ensemble Empirical Mode Decomposition

As proposed by Yeh et al. [34], CEEMD is able to decompose the original sequence to obtain a set of intrinsic mode function (IMF) and trend terms. It has obvious advantages in handling time-frequency sequences of non-linear and non-stationary signals. The components contain the local features of the original sequence at different scales. The algorithm steps are described below.

First, n groups of auxiliary white noise, including positive and negative noise, are added to the original sequence B(t):

$$\left[\begin{array}{c}{H}_1\\ H_2\end{array}\right]=\left[\begin{array}{cc}1& 1\\ 1& -1\end{array}\right]\left[\begin{array}{c}B\\ N\end{array}\right],$$

(2)

where N is an auxiliary sequence, and H₁ and H₂ are positive noise and negative noise sequences, respectively. Then, the obtained sequences are decomposed by EMD to obtain m IMF components, and each set of components is denoted as $C_{ij}^{+}\left(t\right)$ and $C_{ij}^{-}\left(t\right)$, where i = 1, …, n and j = 1, …, m. The value of ${\mathrm{IMF}}_j$ is obtained according to the following formula:

$${\mathrm{IMF}}_j=\frac{1}{2n}{\displaystyle \sum }_{i=1}^n\left(C_{ij}^{+}\left(t\right)+C_{ij}^{-}\left(t\right)\right).$$

(3)

The obtained IMF values are used as the final decomposition results. The original sequence is decomposed into:

$$B_{\left(t\right)}={\displaystyle \sum }_{j=1}^m{\mathrm{IMF}}_{j\left(t\right)}+r_{\left(t\right)}$$

(4)

where $r_{\left(t\right)}$ is a residual trend quantity.

2.2.3. Long Short-Term Memory

As proposed by Hochreiter et al. [35], the LSTM network is a special type of recurrent neural network (RNNs) that is able to learn long-term dependent information in data transmission and can effectively solve the gradient problem. The LSTM network has a more complicated repetition module than RNN, as shown in Figure 2, where X_t is the input sequence at time t, h_t is the output of the LSTM network cells at time t, and σ and tanh are the sigmoid activation function and tanh activation function, respectively. The cell state is the key to this repetitive chain of neural network modules. The cell state is the horizontal line that runs through each module, and it ensures the invariance of the information transfer. It is similar to a conveyor belt that runs through the entire chain. Through the “gate” structure, cell state information is added or removed. The “forgetting gate” determines what information is removed from the cell state. The “input gate” determines what information is stored in the cell state. The “output gate” uses a sigmoid layer to determine the portion of the cell state to be output [22]. After the operation (the circle part of Figure 2), the result is passed to the next unit structure.

2.2.4. Framework of the Hybrid CEEMD-LSTM Model

The workflow of the hybrid model proposed in this paper is shown in Figure 3. The main processes of forecasting through the hybrid model are as follows:

(1) The CEEMD is used to decompose the original SPI sequence into several subsequences and a residual sequence;

(2) The decomposed sequence is predicted using the LSTM network;

(3) Finally, the predicted SPI sequences are obtained by merging the predicted subsequences and residual error.

2.2.5. Evaluation Metrics

In this study, the Kruskal–Wallis test was used to evaluate the accuracy of the estimated values of the two models. When the p-value ≤ 0.05, there is a difference between the estimated and actual values. Four statistical criteria, including root mean square error (RMSE), mean absolute error (MAE), Willmott index (WI), and Nash-Sutcliffe efficiency (NSE), were used to evaluate the performance of the models. The model with the lowest RMSE and MAE as well as the highest WI and NSE was proposed as the appropriate model. The formulas for the above criteria are as follows:

$$RMSE=\sqrt{\frac{1}{N}{\displaystyle \sum }_{i=1}^N{\left(y_i-\widehat{y_i}\right)}^2}$$

(5)

$$MAE=\frac{1}{N}{\displaystyle \sum }_{i=1}^N\left|y_i-\widehat{y_i}\right|$$

(6)

$$WI=\left|1-\left[\frac{{{\displaystyle \sum }}_{i=1}^N{\left(y_i-\widehat{y_i}\right)}^2}{{{\displaystyle \sum }}_{i=1}^N{\left(\left|y_i-\overline{y}\right|+\left|\widehat{y_i}-\overline{y}\right|\right)}^2}\right]\right|$$

(7)

$$NSE=1-\frac{{{\displaystyle \sum }}_{i=1}^N{\left(y_i-\widehat{y_i}\right)}^2}{{{\displaystyle \sum }}_{i=1}^N{\left(y_i-\overline{y}\right)}^2}$$

(8)

where $y_i$ is the observed value, $\overline{y}$ is the average value of $y_i$, $\widehat{y_i}$ is the forecasted value, and $N$ is the total data size of $y_i$.

3. Results and Discussion

The calculation of the SPI and the fitting of the LSTM were both accomplished on the Python 3.7 platform.

3.1. SPI Values at Different Time Scales

The 1-, 3-, 6-, 9-, 12-, and 24-month timescale SPI values were calculated using daily precipitation data from 32 meteorological stations in the Xinjiang Uygur Autonomous Region during the period of 1960–2019. The results of the multi-timescale SPI for the four example stations are shown in Figure 4. It can be seen that the SPI3, SPI6, SPI9, SPI12, and SPI24 values of the four stations show an increasing trend, especially at Fuhai station, and the frequency of extreme drought decreases over time. In the past decade, there were fewer occurrences of extreme drought at the four stations.

Through the Mann–Kendall trend test (

Table 2. Mann–Kendall trend test for SPI sequences.

Example Stations	SPI Series	p Value	Trend
Fuhai	SPI1	2.492 × 10−5	increasing
	SPI3	3.721 × 10−10	increasing
	SPI6	8.860 × 10−14	increasing
	SPI9	1.665 × 10−14	increasing
	SPI12	2.220 × 10−16	increasing
	SPI24	0.000	increasing
Kuerle	SPI1	0.081	no trend
	SPI3	0.001	increasing
	SPI6	3.329 × 10−5	increasing
	SPI9	6.254 × 10−6	increasing
	SPI12	8.926 × 10−7	increasing
	SPI24	4.159 × 10−12	increasing
Yutian	SPI1	0.035	increasing
	SPI3	0.001	increasing
	SPI6	1.649 × 10−4	increasing
	SPI9	2.321 × 10−4	increasing
	SPI12	4.652 × 10−4	increasing
	SPI24	1.990 × 10−11	increasing
Hami	SPI1	0.006	increasing
	SPI3	9.700 × 10−7	increasing
	SPI6	3.212 × 10−12	increasing
	SPI9	1.332 × 10−15	increasing
	SPI12	0.000	increasing
	SPI24	0.000	increasing

), it became clear that all SPI sequences at the four stations have an increasing trend, except for SPI1 at the Kuerle station. The precipitation in Xinjiang has increased significantly since the mid-1980s, which caused the SPI values to increase.

3.2. LSTM Modeling and Prediction

In this study, based on 1-, 3-, 6-, 9-, 12-, and 24-month timescale SPI values, LSTM modeling was performed and the SPI dataset was divided by grid search and cross-validation. A total of 80% of the data (1960–2007) were selected as the training set and 20% (2008–2019) were selected as the test set.

Before the data are input into the LSTM model for prediction, it needs to be normalized to eliminate the effect of data dimensionality differences and to improve the model training speed. The common activation functions for LSTM networks are sigmoid, tanh, and ReLU [36,37,38]. However, as the layers of the neural network deepen, the gradient tends to disappear in the backpropagation when using the sigmoid function, and the stochastic gradient descent converges slowly when using the sigmoid or tanh. Therefore, ReLU was chosen as the activation function in this paper. The batch size was set to 1, which means that the model weights were updated after each sample. The hidden layer was composed of 25 storage units, and the loss function was the mean squared error (MSE). To prevent overfitting of the model training, the early stopping method was used during the simulation. Specifically, as the number of iterations increased, the MSE gradually decreased and the model accuracy improved. When the MSE value started to increase, the training was stopped. At this point, the model accuracy was at its highest. To ensure that the accuracy of the model could be maximized, the number of iterations was set to 300. The prediction results of the LSTM are shown in Figure 5, Figure 6, Figure 7 and Figure 8.

3.3. Hybrid CEEMD-LSTM Model Prediction Results

The data from 1960 to 2007 were used as the training data. Therefore, only this part of the dataset was decomposed by CEEMD. After several modifications and comparisons of the parameters, the standard deviation of the original time-series was finally set to 0.2, the Gaussian white noise logarithm was 100, and the total number of modes (not including the trend) was 8. The multi-scale SPI was decomposed by CEEMD, and eight IMF components and one trend item were obtained. Taking the decomposition of SPI3 at Hami station as an example, the original sequence and the decomposed subsequence are shown in Figure 9. As can be seen in Figure 9, the fluctuation range of the original series is large, while that of the decomposed IMF components is small. Additionally, the fluctuations of the components tend to be smooth as the decomposition proceeds gradually. It is shown that after CEEMD decomposition, the original sequence with strong fluctuations can be decomposed into a set of IMF components with lower fluctuations, which improves the predictability of the sequence.

In the testing phase, the hybrid CEEMD-LSTM model was used to predict the data from 2008 to 2019, and the prediction results for Fuhai, Kuerle, Yutian, and Hami station are shown in Figure 5, Figure 6, Figure 7 and Figure 8, respectively. For all four stations, there is a significant difference between the LSTM predicted value and the observed SPI value at the 1-month timescale. In particular, the LSTM seems to lose its learning ability when forecasting SPI1 at Hami station (Figure 8). At SPI3, SPI6, SPI9, SPI12, and SPI24, the difference between the LSTM predicted and observed value is reduced. In this study, the data volume at SPI1 was the largest and the time-series was the least smooth, thus, the prediction was most ineffective at that timescale. The SPI time-series is a non-stationary series, but the accuracy of the single model prediction results is strongly influenced by the stationarity of the original data. Similarly, Adikari et al. [39] predicted the SPI using LSTM, and it provided only merely acceptable results at Nonghine, Sekong River Basin.

The predicted value of the hybrid CEEMD-LSTM model is closer to the observed SPI, and the predicted SPI trend is consistent with the actual trend. In Figure 5, Figure 6, Figure 7 and Figure 8, the hybrid model performs well at SPI6, SPI9, SPI12, and SPI24. As can be seen from Figure 6, an extreme drought occurred in the Xinjiang Uygur Autonomous Region in 2011. For this year, the hybrid model’s prediction of drought conditions was more consistent with the actual situation and it more accurately predicted the occurrence of extreme drought. Through CEEMD decomposition, local features of the original series at different scales are extracted, and the non-stationary time-series are transformed into stationary components. This reduces the non-stationarity of SPI sequences, which improves the ability of LSTM to predict SPI sequences. Therefore, the hybrid CEEMD-LSTM model has a good performance, even in forecasting SPI1 at each station.

With an increasing timescale, the accuracy of the predictions of the LSTM and CEEMD-LSTM models improves. At the 24-month scale, the prediction results of both models are nearly consistent with the observed SPI. Since the prediction accuracy of the single model gradually improves, that of the hybrid model also improves. At SPI1, the prediction of the hybrid CEEMD-LSTM model is significantly better than that of the single model because the predictability of the subseries, which are obtained from the CEEMD decomposition, is higher than that of the original series. However, this advantage of the combined model diminishes with an increasing timescale. At SPI12 and SPI24, the CEEMD-LSTM model only slightly outperforms the single model. The reason for this is that the long timescale SPI series aggregates more information from the original data and the whole series tends to be stable. Therefore, the prediction accuracy of the single model is improved. In recent years, several scholars have used hybrid models for forecasting. They have found that the results of hybrid models, especially those combining decomposition methods with forecasting methods, are better than those of single models [22,27,40,41,42]. The difference between the aforementioned studies and the current study is that the current study uses such a hybrid model in drought prediction, and the decomposition method chosen was CEEMD. CEEMD is more suitable for prediction because it can effectively solve the problem of mode aliasing and reduce the residual white noise.

The accuracy of the predicted values of the LSTM and CEEMD-LSTM models was evaluated using the Kruskal–Wallis test. The results showed that the p-values of the predicted sequences were greater than 0.05, except for the LSTM predicted SPI1 values at Fuhai station (0.015) and Hami station (0.003). This means that the differences between the observed values and the predicted values of the hybrid CEEMD-LSTM model is not statistically significant, and the CEEMD-LSTM model is able to predict all timescales at different sites, while the LSTM model may not be proficient in predicting the most unstable SPI1 sequences. Four evaluation metrics, namely MAE, RMSE, NSE, and WI, were used to evaluate the prediction results of the two models (

Table 3. Comparison of the observed and predicted values of the two models using statistical criteria.

Example Stations	SPI Series	Model	MAE	RMSE	NSE	WI
Fuhai	SPI1	LSTM	0.790	1.028	−32.371	0.255
		CEEMD-LSTM	0.675	0.815	−0.633	0.721
	SPI3	LSTM	0.551	0.681	−0.112	0.781
		CEEMD-LSTM	0.474	0.578	0.247	0.850
	SPI6	LSTM	0.389	0.491	0.498	0.885
		CEEMD-LSTM	0.295	0.378	0.714	0.934
	SPI9	LSTM	0.275	0.360	0.682	0.924
		CEEMD-LSTM	0.219	0.291	0.790	0.951
	SPI12	LSTM	0.219	0.294	0.773	0.946
		CEEMD-LSTM	0.169	0.219	0.873	0.970
	SPI24	LSTM	0.152	0.198	0.836	0.960
		CEEMD-LSTM	0.119	0.152	0.895	0.976
Kuerle	SPI1	LSTM	0.648	0.868	−46.280	0.168
		CEEMD-LSTM	0.568	0.739	−0.904	0.688
	SPI3	LSTM	0.604	0.791	−0.217	0.778
		CEEMD-LSTM	0.502	0.674	0.214	0.850
	SPI6	LSTM	0.471	0.660	0.451	0.890
		CEEMD-LSTM	0.412	0.540	0.717	0.934
	SPI9	LSTM	0.386	0.573	0.670	0.933
		CEEMD-LSTM	0.284	0.408	0.866	0.969
	SPI12	LSTM	0.308	0.486	0.794	0.957
		CEEMD-LSTM	0.225	0.342	0.915	0.981
	SPI24	LSTM	0.270	0.427	0.842	0.967
		CEEMD-LSTM	0.196	0.312	0.930	0.984
Yutian	SPI1	LSTM	0.541	0.711	−28.523	0.307
		CEEMD-LSTM	0.509	0.615	−1.257	0.678
	SPI3	LSTM	0.525	0.692	−0.184	0.796
		CEEMD-LSTM	0.486	0.612	0.117	0.842
	SPI6	LSTM	0.419	0.560	0.445	0.889
		CEEMD-LSTM	0.324	0.439	0.740	0.942
	SPI9	LSTM	0.368	0.541	0.513	0.903
		CEEMD-LSTM	0.296	0.418	0.777	0.949
	SPI12	LSTM	0.239	0.393	0.805	0.956
		CEEMD-LSTM	0.196	0.300	0.889	0.975
	SPI24	LSTM	0.178	0.323	0.863	0.967
		CEEMD-LSTM	0.177	0.260	0.908	0.979
Hami	SPI1	LSTM	0.661	0.806	−194.885	0.115
		CEEMD-LSTM	0.517	0.650	−0.614	0.724
	SPI3	LSTM	0.532	0.673	−0.208	0.787
		CEEMD-LSTM	0.470	0.583	0.231	0.851
	SPI6	LSTM	0.397	0.574	0.479	0.884
		CEEMD-LSTM	0.352	0.466	0.654	0.924
	SPI9	LSTM	0.342	0.501	0.578	0.904
		CEEMD-LSTM	0.286	0.406	0.722	0.937
	SPI12	LSTM	0.263	0.420	0.687	0.925
		CEEMD-LSTM	0.216	0.315	0.819	0.958
	SPI24	LSTM	0.159	0.244	0.790	0.950
		CEEMD-LSTM	0.139	0.201	0.852	0.966

). At the four stations, the MAE values of LSTM were above 0.5 at SPI1, and were mostly below 0.2 at SPI24. The MAE value tends to decrease with an increasing timescale, and RMSE displays the same, while NSE and WI display the opposite trend. These trends indicate that the prediction accuracy of LSTM constantly improves with an increasing timescale. The hybrid CEEMD-LSTM model outperforms the LSTM model in prediction at different timescales, which means that CEEMD can effectively improve the prediction accuracy of the LSTM model. At SPI24, the NSE values of the hybrid model for the four sites were 0.895, 0.930, 0.908, and 0.852, respectively, and the WI values were all above 0.95.

As seen in Figure 6, extreme drought has occurred in the Xinjiang Uygur Autonomous Region during the past 10 years, specifically in 2011. The LSTM and CEEMD-LSTM model prediction results of 32 meteorological stations in 2011 were visualized by the Empirical Bayesian Kriging (EBK) in ArcGIS, as shown in Figure 10. From the figure, we can see that the prediction of the hybrid CEEMD-LSTM model for the spatial distribution of drought is closest to the actual situation. Based on the advantage of CEEMD in non-stationary signal processing, the hybrid model has better performance in SPI prediction.

4. Conclusions

In this study, the LSTM and CEEMD-LSTM models were used to forecast SPI at six timescales in the Xinjiang Uygur Autonomous Region, China. Four statistical criteria were selected to evaluate the prediction accuracy of the two models, and the spatial distribution of the prediction results of the LSTM and CEEMD-LSTM models were visualized using the EBK method.

Based on the values of the statistical criteria (Table 3), the performance of both models increased gradually with increasing timescale. They were lowest at the 1-month scale and highest at the 24-month scale. Moreover, the prediction accuracy of the hybrid model was higher than that of the single model at different timescales. At SPI24, the NSE values of the hybrid model were all greater than 0.85, which shows that CEEMD has advantages in dealing with non-stationary and non-linear data. Through CEEMD decomposition, the original time sequence becomes stable and the predictability of the sequence is improved. Based on Figure 10, we can see that the spatial distribution of drought predicted by the hybrid CEEMD-LSTM model is more consistent with the actual situation. As the prediction accuracy of the hybrid model is better than that of the single model at each timescale, the hybrid model composed of CEEMD and LSTM can contribute to improving the prediction accuracy of meteorological drought.

There are some limitations to this study. In this study, the SPI was chosen to reflect the drought conditions based on its multi-timescale application characteristics and can be calculated by precipitation alone. However, in recent years, both the precipitation and temperature in the Xinjiang Uygur Autonomous Region have shown an increasing trend, and thus the drought characteristics cannot be fully reflected by precipitation. In the subsequent study on the applicability of the combined model, the prediction of several drought indices, such as the SPEI, should be considered.