Hi Sampson,
From what I understand, the project involves training the DQN algorithm to track the PV system's maximum power under standard testing conditions. However, when the irradiance drops below 800W/m2, the power level observed falls below the expected level.
When the irradiance drops below 800W/m2, it is expected that the power level will decrease. However, if the power level falls significantly lower than expected, there could be several factors contributing to this issue.
There are a few approaches you can consider:
- Insufficient training data: The DQN agent might not have enough training data for low irradiance conditions. Ensure that the training dataset includes enough samples covering a wide range of irradiance levels, including low irradiance scenarios.
- Exploration-exploitation trade-off: The DQN agent might not be exploring enough during training to discover optimal policies for low irradiance conditions. Adjusting the exploration rate or using exploration strategies like epsilon-greedy or Boltzmann exploration.
- Limited action space: The action space of the DQN agent may not be suitable for handling low irradiance conditions. Consider expanding the action space to allow for more fine-grained control over the PV system parameters, such as adjusting the voltage or current levels.
Attached below are some documentation links that you may find helpful:
- Get Started with Reinforcement Learning Toolbox (mathworks.com)
- Deep Learning Toolbox Documentation (mathworks.com)
Hope this helps!
