In this paper, an ensemble learning approach is proposed for load forecasting in urban power systems. The proposed framework consists of two levels of learners that integrate clustering, Long Short-Term Memory (LSTM), and a Fully Connected Cascade (FCC) neural network. Historical load data is first partitioned by a clustering algorithm to train multiple LSTM models in the level-one learner, and then the FCC model in the second level is used to fuse the multiple level-one models. A modified Levenberg-Marquardt (LM) algorithm is used to train the FCC model for fast and stable convergence. The proposed framework is tested with two public datasets for short-term and mid-term forecasting at the system, zone and client levels. The evaluation using real-world datasets demonstrates the superior performance of the proposed model over several state-of-the-art schemes. For the ISO-NE Dataset for Years 2010 and 2011, an average reduction in mean absolute percentage error (MAPE) of 10.17% and 11.67% are achieved over the four baseline schemes, respectively.