- Machine learning techniques support better solar power plant forecasting.
- Machine learning techniques play a crucial role in deciding where to build a plant when accurate or limited location data is available.
- Machine learning techniques help maintain smart grid stability.
The global solar photovoltaic (PV) installed capacity in 2013 was 138.9 GW and it is expected to grow to over 455 GW by 2020. However, solar power plants still have a number of limitations that prevent it from being used on a larger scale. One limitation is that the power generation cannot be fully controlled or planned for in advance since the energy output from solar power plants is variable and prone to fluctuations dependent on the intensity of solar radiation, cloud cover and other factors. Another important limitation is that solar energy is only available during the day and batteries are still not an economically viable storage option making careful management of energy generation necessary. Additionally, as the installed capacity of solar power plants grows and plants are increasingly installed at remote locations where location data is not readily available, it is becoming necessary to determine their optimal sizes, locations and configurations using other methods. Machine learning techniques provide solutions that have been more successful in addressing these challenges than manually developed specialized models.
Accurate forecasts of solar power production are a necessary factor in making the renewable energy technology a cost-effective and viable energy source. Machine learning techniques can correctly forecast solar power plant generation at a better rate than current specialized solar forecasting methods. In a study conducted by Sharma et al, multiple regression techniques including least-square support vector machines (SVM) using multiple kernel functions were used in the comparison with other models to develop a site specific prediction model for solar power generation based on weather parameters. Experimental results showed that the SVM model outperformed the others with up to 27 percent more accuracy.
Furthermore, machine learning techniques play a crucial role in assisting decision making steps regarding the plants location selection and orientation selection as solar panels need to be faced according to solar irradiation to absorb the optimal energy. Conventional methods for sizing PV plants have generally been used for locations where the required weather data (irradiation, temperature, etc.) and other information concerning the site is readily available. However, these methods cannot be used for sizing PV systems in remote areas where the required data are not available, and thus machine learning techniques are needed to be employed for estimation purposes. In a study conducted by Mellit et al., an artificial neural network (ANN) model was developed for estimating sizing parameters of stand-alone PV systems. In this model, the inputs are the latitude and longitude of the site, while the outputs are two hybrid-sizing parameters. In the proposed model, the relative error with respect to actual data does not exceed 6 percent, thus providing accurate predictions. This model has been evaluated on 16 different sites and experimental results indicated that prediction error ranges from 3.75-5.95 percent with respect to the sizing parameters. Additionally, metaheuristic search algorithms address plan location optimization problems by providing improved local searches under the assumption of a geometric pattern for the field.
Lastly, to maintain grid stability, it is necessary to forecast both short term and medium term demand for a power grid with renewable energy sources contributing a considerable proportion of energy supply. The MIRABEL system offers forecasting models which target flexibilities in energy supply and demand, to help manage the production and consumption in the smart grid. The forecasting model combines widely adopted algorithms like SVM and ensemble learners. The forecasting model can also efficiently process new energy measurements to detect changes in the upcoming energy production or consumption. It also employs different models for different time scales in order to better manage the demand and supply depending on the time domain.
Ultimately, machine learning techniques support better operations and management of solar power plants.