•A multi-time-scale approach is proposed for electric power demand forecasting.•Historical load is modeled as a time-series with multiple seasonality levels.•Each seasonal cycle of load data is studied without additional input.•Akaike/Bayesian information criteria is used for accuracy evaluation.•Box-Jenkins method is enhanced for modeling the load data over a time period. Short-term load forecasting is essential for reliable and economic operation of power systems. Short-term forecasting covers a range of predictions from a fraction of an hour-ahead to a day-ahead forecasting. An accurate load forecast results in establishing appropriate operational practices and bidding strategies, as well as scheduling adequate energy transactions. This paper presents a generalized technique for modeling historical load data in the form of time-series with different cycles of seasonality (e.g., daily, weekly, quarterly, annually) in a given power network. The proposed method separately models both non-seasonal and seasonal cycles of the load data using auto-regressive (AR) and moving-average (MA) components, which only rely on historical load data without requiring any additional inputs such as historical weather data (which might not be available in most cases). The accuracy of data modeling is examined using the Akaike/Bayesian information criteria (AIC/BIC) which are two effective quantification methods for evaluation of data forecasting. In order to validate the effectiveness and accuracy of the proposed forecaster, we use the hourly-metered load data of PJM network as a real-world input dataset.