Aiming at the output prediction of small hydropower units, a prediction method of small hydropower based on the improved gray wolf algorithm and adaptive similar day selection is proposed. Firstly, the load data was divided by lunar calendar according to the output law of small hydropower. Considering that the influence of various factors on the output of small hydropower was variable, the adaptive similar day selection method and an optimization strategy for the weights of each impact factor based on the improved grey wolf optimization (GWO) algorithm were employed. Then, the selected similar daily samples were fed into RBF and BP network for small hydropower units’ output forecasting, respectively. Later, the above predicted results were input into the generalized regression neural network(GRNN) optimized by the grey wolf optimization for nonlinear combination forecasting. By the analysis of an example in a certain area, the mean absolute error of the proposed prediction model is reduced by 3.28%, 1.73% and 0.29% respectively, compared with the BP, RBF and non-optimized GRNN combined prediction model, which verifies the validity of the proposed model.