As a non-linear, non-stationary and fluctuating time series, daily peak load is difficult to predict accurately. In this paper, a gated recurrent neural network (GRNN) combined with dynamic time warping (DTW) is proposed to predict daily peak load accurately. DTW distance is used to match the most similar load curve, which can capture the trend of load change. The thermal coding scheme is used to encode the discrete variables and extend their characteristics to represent the influence on the load curve. A dtw-gru algorithm based on DTW is proposed for daily peak load forecasting, and it is tested on the European Intelligent Technology Network (EUNITE) dataset. Simulation results show that the MAPE of this algorithm is only 1.01% compared with other algorithms.