In electricity demand forecasting, a common problem is the presence of seasonal patterns. The infeed data exhibits daily, weekly and yearly seasonal patterns. We design a multi-step approach to overcome this problem. Firstly, a separate model is introduced for every hour of a day so as to remove the daily cycle of the infeed pattern similar to Soares and Medeiros [18]. Each model becomes much simpler by avoiding the intra-day patterns. Following this, we include the average hourly infeed of the last three years into the model to simulate the yearly season. This variable reflects the assumption that the future electricity demand follows the same pattern as historical demand, and the future demand is thus correlated with the moving average of historical values.