How were the forecasts for optimizing demand side generated?

Hi @bull,

Would you be able to tell us how the load and PV forecasts were generated for each site?


That’s an interesting question. My guess is that they simply add random noise to true values.

It tried to improve the forecasts using machine learning but I only could improve the accuracy about a 2%.

1 Like

I’m doing the master thesis on this topic in Lappeenranta University of Technology. Our team used 6 heuristics and reinforcement learning. We picked the best model for the each site based on the score on previous runs and time-series statistics. It gave us about -0.15xx.

@hoa.nguyen,, @Barbe, @ironbar could you share your approach in this task please? Your scores are much higher than our. What is your secret?)

1 Like

Hi Denis,
If I end on the price positions I will probably make a detailed post of my solution.

The summary of my solution:

  • Dynamic programming
  • Simplification of the problem. After thinking I discovered that I could group the timesteps into bigger blocks. The reason for making divisions were: changes on prices and changes on the balance of the system(positive balance or negative balance). So instead of optimizing for 96 timesteps the problem was simplified to a few blocks.
  • Simplification of the actions available at each timestep. I carefully think of which actions had sense depending on the state of the battery and the balance of the system. This was important to do a fast optimization.
  • Small improvement of the forecast using linear regression. But the effect was very small.