Power generation using renewable energy such as wind and solar energy depend on weather behavior. An accurate weather forecast improves the operation management of a Smart Grid. However, the environmental conditions vary with both time and space and, depending on the location have different classifications. The environmental conditions vary with time and depending on the country, also the changes between standard, local and summer time can result in days of 23 h, to obtain days of 24 h; the two adjacent hours are used to calculate the missing hour. In this paper is addressed a nonlinear symbolic regression (RS by its acronym in Spanish) model is used to make annual demand forecasts. This algorithm is assessment with real measurements creating models that conform to the load forecasting of 28 weather stations. The learning variables are loads of demand and temperatures; that are grouped by hours for a two-year time interval. The RS-DEEPSO algorithm is proposed to improve the forecast. This is used due to its evolutionary, self-adaptive qualities and the possibility of exploring new scenarios in search of optimum. The forecast obtained using the new hybrid RS-DEEPSO has an average error per day less than 3.9% and per hour is less than 5.62% during a period of 1 year.