Deep learning approaches have shown superior advantages than shallow techniques in the field of electrical load forecasting; however, their applications in existing studies encounter thorny issues despite their excellent forecasting performance: heavy computing costs due to complicated network structure and restricted to the deterministic point forecasting. This paper aims to solve above two problems by proposing a deep learning approach based on K-nearest neighbors to capture uncertainty and reflect the range of electrical load fluctuation. First, the K-nearest neighbors algorithm is applied to seek features of historical electrical load time series that are similar to the future values by calculating the distance between the training and testing datasets. Then the second generation of non-dominated sorting genetic algorithm is adopted for multi-objective optimization to find out the smallest category number of K-nearest neighbors and the highest forecasting accuracy. Based on the forecasting results of the deep belief network, modified non-parameter kernel density estimation is used to obtain the prediction intervals. Five datasets collected from Australia are employed to examine the effectiveness of the proposed model. By a series of comparisons with other state-of-the-art models, experimental results confirm that the proposed interval forecasting model cannot only improve the forecasting efficiency and accuracy, but also simplify the forecasting process of deep learning approaches, which can provide great referential value for future work.