Abstract: Aimed at the load time series data has linear and nonlinear complex characteristics in the data center, a single model often shows certain limitations in capturing the characteristics and forecasting. In this regard, a method combined of ARIMA and IndRNN and incorporating wavelet decomposition is proposed. The series was decomposed into trend subsequences and detail subsequences by Haar wavelet. The ARIMA and IndRNN models were respectively used to model and predict the two subsequences, and reconstructed the two prediction to get the combined model’s first prediction results. The error series was predicted by the IndRNN model which was to further improve the prediction accuracy. The results show that the combined model of ARIAM and IndRNN is reliable and has higher accuracy than other methods.