I am using a hidden Markov model (HMM) to forecast day-ahead (hourly) solar irradiance, which is a measure of the intensity of the sun. As a part of my project, I am constrained to use only the past irradiance data and not any external variables that can affect the sun's intensity (like cloud, temperauture, etc.). I have generated 3 states from the HMM - high, medium, and low, to represent the daily irradiance pattern.
For prediction, I have calculated the weights for each of the states using the emission probability and the probability transition matrix. Now, I have tested and verified that choosing the best state using the weights calculated would give me a rough forecast with fair accuracy. What I am trying to do next is capture the trend better by creating sort of a band around the highest weighted state, where my forecast values can move around to give better prediction. Can anyone please suggest me a method of doing this effectively?
Dear Abhilasha
I suggest you train the hidden state using Baum-Welch algorithm and then you can use Viterbi Algorithm for finding best state.
Regards
Toto
Hi Toto Haryanto,
Thank you for your response! Are you suggesting that I train again the hidden states (high, medium, and low) that I got from the HMM with the Baum-Welch algorithm? If so, is there some way I can incorporate some recent irradiance data (maybe 4 days ahead of the target day) to capture sudden atmospheric changes like cloud or rain?
Thanks!
Dear Abhilasha
Thank for your response also. In my experience, Baum-Welch usually used for continue dataset such as signal of voice, rainfall and etc. I'm s Sorry because I not clear with your data type. If type of your data is discrete, may be you can compute the probability for each state over all possible state. I just have little-bit experience using hidden semi markov for may thesis project for protein sequence analysis.
Regards,
Toto
- Badges
- Science topic
- Similar topics
- Control Systems Engineering
- Filtering