The forecast and estimation of total electric power demand of a residential community, its baseload, and its heating ventilation and air-conditioning (HVAC) power component, which represents a very large portion of a community electricity usage, are important enablers for optimal energy controls and utility planning. This paper proposes a method that employs machine learning in a multi-step integrated approach. An LSTM model for total electric power at the main circuit feeder is trained using historic multi-year hourly data, outdoor temperature, and solar irradiance. New key temperature indicators, TmHAVC, corresponding to the standby zero-power operation for HVAC systems for summer cooling and winter heating are introduced using a V-shaped hourly total load curve. The trained LTSM model is additionally run with TmHVAC and zero irradiance inputs yielding an estimated baseload, which is representative of typical occupancy patterns. The HVAC power component is disaggregated as the difference between total and baseload power. Total power forecasts of an aggregated residential community as seen by major distribution lines are experimentally validated with a satisfactory MAPE error below 10% based on a 4-year dataset from a representative suburban community with more than 1800 homes in Kentucky, U.S. Discussions regarding the validity of the separation method based on combined considerations of fundamental physics, statistics, and human behavior are also included.