Skip to main content

T-00371: PV Power Forecasting in Remote Microgrids based on the Markov Switching Model


The utilization of Photovoltaic (PV) systems is increasing rapidly in diesel based remote microgrids to reduce diesel fuel consumption which contributes to the primary cost of microgrid operation. However, load does not always correlate with the PV power availability. In addition, PV systems have challenges of uncertainty and variation. Due to the variation, a large reserve is required to compensate fluctuation and improve reliability. This causes an increment in the operational cost as the generators are now forced to operate in a lower efficiency region to provide the required reserve margin. Solar irradiance forecasting and PV power forecasting would greatly reduce the reserve requirement and improve PV energy utilization. Moreover, with PV forecast, energy management system can better manage dispatchable resources and improve energy efficiency. Most of the solar irradiance forecasting techniques require information like weather and satellite data which may not be readily available in remote areas or would require additional sensors. A new solar irradiance forecasting method that makes use of locally available data is proposed. This technique is well suited for remote microgrids, since it does not require any additional data besides historical solar irradiance data to forecast PV irradiance.


This solar irradiance forecasting method based on the Markov Switching Model considers past solar irradiance data, clear sky irradiance and Fourier basis functions to create linear models for three regimes or states: high, medium and low energy regimes for a day corresponding to sunny, mildly cloudy and extremely cloudy days, respectively. Fourier basis functions are used for representing daily and yearly periodic variation of solar irradiance. The best model is selected using Bayesian Information Criterion (BIC). A case study for Brookings, SD, resulted in an average Mean Absolute Percentage Error (MAPE) of 31.8% for 2001 to 2005 with higher errors during summer months than during winter months. Similarly a case studied for the same location resulted in Root Mean Square Error (RMSE) of 79.4 W/m2 and MAPE of 21.7 % for July 24, 2012.


  • This solar irradiance forecasting method simply uses publicly and freely available historical solar irradiance data for forecasting irradiance for a location.
  • This method does not need information like weather and satellite data.
  • This method is suitable for remote locations which do not have any communication infrastructure.