Course project to predict solar radiation data using and RBF kernel SVM comparing its performance with a Neural Network. It is an implementation of Short Term Solar Power Prediction
The project takes current radiation data and weather conditions to predict the radiation for a particlar window (has to be decided beforehand). Later experimentation can involve using ensemble models to predict for multiple windows and to increase the performance.
We acquired the last 15 years hourly Solar Radiation Data from the NSRDB Database. All the data used here is that for Mumbai from 2000-2015.
The plots above show the two visualizations used for the dataset, on the left is a one-dimensional plot of the radiation, and on the right, there is a heatmap. The 1D plot is useful for seeing trends in seasons and the 2D plot is helpful for seeing trends in daily radiation values.
The results for Mean Average Error (MAE) and Mean Average Percentage Error have been plotted below for different prediction windows.
We experimented by changing the length of the dataset to see how it performs. The results of those experiments are plotted.
.png)
We tested the SVM model against the other model and found it to be superior in all prediction windows.