The non-intrusive load decomposition is to decompose the power signal of a single load device according to the known total power signal. At present, there are many problems based on deep learning models, such as insufficient load feature extraction, low decomposition accuracy, large decomposition error of load equipment with low frequency, and so on. To solve these problems, this paper proposes an attention recurrent neural networks (ARNN) model to realize non-invasive load decomposition, which combines regression network and classification network to solve the non-invasive load decomposition problem. The model extracts the features of sequence signals through RNN network, and uses the attention mechanism to locate the position of important information in the input sequence, so as to improve the representation ability of network. Experiments on public datasets Wiki-Energy and UK-DALE show that our proposed deep neural network is superior to the most advanced neural network under all experimental conditions. We also show that the attention mechanism and auxiliary classification network can correctly detect the opening or closing of devices, and locate the high-power signal part, which improves the accuracy of load decomposition.