Adjusted combination of moving averages: A forecasting system for medium-term solar irradiance

Abstract Global Horizontal Irradiation forecasts are necessary for an efficient use of fluctuating energy output from photovoltaic plants. The purpose of this paper is to provide efficient, easy-to-implement Global Horizontal Irradiation forecasts on an hourly basis for a medium-term horizon (longer than 48 h). These forecasts are essential for the strategic deployment of tasks including unit commitment, transmission management, trading, hedging, planning, asset optimization, maintenance scheduling and spinning of power units. Nonetheless, forecasting models for medium term horizons are scarce. This work intends to bridge that gap by proposing a method based on an adjusted combination of moving averages which includes the yearly cycle in addition to the day–night cycle. This method is straightforward to implement and easy to integrate into corporate computing systems. We compare our approach with several well-known alternative methods, such as deep recurrent neural networks or autoregressive integrated moving average models, among others. The results show that recurrent neural networks are 10% more accurate than the second best method, and 90% better than our proposal for very short horizons (up to 5 h ahead). However, this advantage is lost when longer horizons are tested, especially for horizons longer than 30 h ahead, for which our method produces forecasts which are 20% more accurate than those made by recurrent neural networks. The reason behind that improvement at longer horizons is the inclusion of the yearly cycle in the proposed approach.