Long Short-Term Memory (LSTM) neural network is an enhanced Recurrent Neural Network (RNN) that has gained significant attention in recent years. It solved the vanishing and exploding gradient problems that a standard RNN has and was successfully applied to a variety of time-series forecasting problems. In power systems, distribution feeder long-term load forecast is a critical task many electric utility companies perform on an annual basis. The goal of this task is to forecast the load change on existing distribution feeders for the next few years. The forecasted results will be used as input in long-term system planning studies to determine necessary system upgrades so that the distribution system can continue to operate reliably during normal operation and contingences. This research proposed a comprehensive hybrid model based on LSTM neural network for this classic and important forecasting task. It is not only able to combine the advantages of top-down and bottom-up forecasting models but also able to leverage the time-series characteristics of multi-year data. This paper firstly explains the concept of LSTM neural network and then discusses the steps of feature selection, feature engineering and model establishment in detail. In the end, a real-world application example for a large urban grid in West Canada is provided. The results are compared to other models such as bottom-up, ARIMA and ANN. The proposed model demonstrates superior performance and great practicality for forecasting long-term peak demand for distribution feeders.