Abstract: With the increase of renewable energy penetration, the source-load balance and stable operation of power systems depend on more accurate and reliable forecasts. The net load is the actual load minus the renewable energy generation, and its accurate prediction can effectively improve the economy and safety of the power system. Therefore, this paper adopts a direct prediction strategy and proposes a net load prediction model based on a self-attention encoder and deep neural network, including a self-attention encoder module that extracts the original uncertainty feature information and a long and short-term memory neural network module that extracts the net load temporal features, and these two parts of feature information are input into a residual neural network to output the final prediction results. At the same time, since the net load integrates several uncertainties such as load, PV, wind power, and it is highly volatile, this paper combines conditional quantile regression to effectively implement non-parametric interval prediction to quantify forecast uncertainty and evaluate the range of net load fluctuations. Case studies show that the proposed AE-DNN forecasting model achieves higher net load forecasting accuracy than common forecasting models, and the quality of the prediction intervals is better than that of the baseline model, which can effectively support the real-time grid operation.