In Taiwan, even though the average annual rainfall is up to 2500 mm, water shortage during the dry season happens sometimes. Especially in recent years, water shortage has seriously affected the agriculture, industry, commerce, and even the essential daily water use. Under the threat of climate change in the future, efficient use of water resources becomes even more challenging. For a comparative study, support vector machine (SVM) and other three models (artificial neural networks, maximum likelihood classifier, Bayesian classifier) were established to predict reservoir drought status in next 10-90 days in Tsengwen Reservoir. (The ten-days time interval was applied in this study as it is the conventional time unit for reservoir operation.) Four features (which are easily obtainable in most reservoir offices), including reservoir storage capacity, inflows, critical limit of operation rule curves, and the number of ten-days in a year, were used as input data to predict drought. The records of years from 1975 to 1999 were selected as training data, and those of years from 2000 to 2010 were selected as testing data. The empirical results showed that SVM outperforms the other three approaches for drought prediction. Unsurprisingly the longer the prediction time period is, the lower the prediction accuracy is. However, the accuracy of predicting next 50 days is about 85% both in training and testing data set by SVM. As a result, we believe that the SVM model has high potential for predicting reservoir drought due to its high prediction accuracy and simple input data.

The support vector machine (SVM) has been applied to drought prediction and it typically yields good performance on overall accuracy. However, the prediction accuracy of the drought category is much lower than that of the non-drought and severe drought categories. In this study, a two-stage approach was used to improve the SVM to increase the drought prediction accuracy. Four features, (1) reservoir storage, (2) inflows, (3) critical limit of operation rule curves, and (4) the Nth ten-day in a year, were used as input data to predict reservoir drought. We used these features as input data because they are the most commonly kept records in all reservoir offices. Empirical results show that the two-stage SVM outperforms the original SVM and the three other approaches (artificial neural networks, maximum likelihood classifier, Bayes classifier) for drought prediction. Not surprisingly, the longer the prediction time period, the lower the prediction accuracy is. However, the accuracy of predicting conditions within the next 50 days was approximately 85% both in training and testing data set by the two-stage SVM. Drought prediction provides information for reservoir operation and decision making in terms of water allocation and water quality issues. The result shows the benefit of a two-stage approach of SVM for drought prediction, as the accuracy of drought prediction increased quite substantially.