Forecasting solar power production

Why do we need to forecast solar power production?

Knowing how much solar power will be produced in the future is useful for many reasons. How we use these forecasts depends on how far ahead we are looking (the forecasting horizon):

• Estimates of long-term changes in solar power production because of climate change can be used for policy decisions and energy system planning.
• Predicting solar power for the next two weeks helps schedule maintenance and avoids losing power when equipment is down.
  
• Forecasts 1-2 days ahead can inform bids in the day-ahead energy market or assist in preparing backup energy for the grid, helping to keep the energy system stable. 
• Forecasts 1-6 h ahead help with timely activation of reserve energy, correcting imbalances in the energy system.
• Sub-hourly forecasting horizons improve grid stability and solar power plant operation, e.g. through better coordination of assets in reserve markets, or better operation of solar power combined with batteries.

How can we generate solar power forecasts?

When forecasting solar power, future irradiance and temperature are most important. Although future system performance is also an important factor, its predictability is much higher than the weather. Cloud dynamics are particularly challenging to forecast, and efforts to improve solar power forecasting are therefore often focused on clouds and irradiance. For irradiance forecasts, the forecasting horizon decides how the forecasts are generated. At horizons above 6 h, it's best to use weather models for the short-term, and climate models for the long-term. For forecasts between 1 and 6 hours, analyzing cloud movement in satellite images is effective. Sub-hourly forecasts need localized irradiance information, e.g., from sky imagers, irradiance sensors, or solar power systems. In the past, forecasting approaches have been broken down into empirical descriptions of separate but connected physical phenomena. Recently, machine learning approaches that handle the problem as a whole, have proven to be effective, e.g., ECMWF’s AIFS as seen in Fig 1

Solar power forecasting in FME SOLAR

FME SOLAR conducts research in solar power forecasting using both traditional and machine learning approaches, and at several forecasting horizons:

• Climate change influence on solar energy’s role in the energy system.
• Probabilistic day-ahead forecasting through novel machine learning algorithms.
• How forecast uncertainty information can help optimize micro-grid control.
• At sub-hourly horizons, we explore machine learning on decomposed solar power time series, deep learning applied to sky images (Fig 2), and extraction of cloud movement from production data.