Reservoir computing is an effective machine learning paradigm. We propose a new reservoir computing called delayed deep reservoir computer (DDRC), DDRC is based on feedforward neural networks (FNNs), the input of the current time and delayed time are projected into the high-dimensional space by FNNs in parallel, and are combined into a high-dimensional feature representation through a set of weight coefficients. Feature representation vectors are mapped to the output by linear regression. DDRC retains the advantages of simple training and small computation of ESN, only readout weight requires learning from the data. Several numerical experiments show that DDRC achieves better results in chaotic time series prediction. In particular, DDRC enables long term prediction of multiple superimposed oscillator (MSO) sequence.