The invention discloses an ARIMA-SVR-based hydrological time series abnormal-value detection method. The method includes the following steps: firstly, acquiring hydrological time series data, and carrying out ARIMA fitting on a hydrological time series, wherein an ARIMA model can well fit a linear part of the data, but a fitting effect is poorer when a nonlinear part exists in the data; then obtaining residual errors of an ARIMA part, and using a method of 10-fold cross-validation to find out optimum combination of gamma, cost and a kernel function of SVR; and finally, adding fitting values oftwo parts together to obtain final prediction values, obtaining a confidence interval, of which confidence is p, by solving, comparing the prediction values with the confidence interval, and determining that values outside the confidence interval are abnormal values. The invention provides, for practitioners related to water conservancy, a method of finding abnormal values in hydrological time series, and detection of the abnormal values in the hydrological time series has important guiding significance for work of flood control, drought relief and the like.