Daily streamflow prediction is of considerable value for society, particularly for agricultural and hydrological management. With substantial progress on machine learning (ML) having been made over the past decade, streamflow prediction has emerged as a potential target area for different kinds of neural networks. However, most ML models still require empirical separation of baseflow, or for the predictors to contain historical streamflow data, making future projection with these models difficult. In our research, we used a dilated convolutional neural network to predict streamflow in California. The predictors, which are precipitation, relative humidity, downward shortwave flux and temperature, are fed into the network in batches using windowing and output streamflow is predicted in the same batch. Using the input window, we can extract temporal features from predictor vectors. The output window can generalize the model to different gauge stations with the same architecture. Compared with hydrological models, the ML model can achieve similar or even higher prediction accuracy. For certain unmanaged stations, Nash-Sutcliffe model efficiency coefficient can achieve values over 0.6. Compared with linear models, the results demonstrate that our model can extract nonlinear, time dependent features to predict streamflow, and generate future projections without historical streamflow as input.