This model was trained using a reinforcement learning (RL) method to control home batteries. We trained this global RL model
on 90 diverse Belgian households with the objective of reducing daily energy cost. We proposed a transfer learning framework
to finetune this pretrained global model on limited data from a target household that you would like to use this model for.
To read more about our transfer learning framework, data, and RL-based method we encourage you to follow our research published in BuildSys 2024.
The model is a neural network which gets 6 arguments as an input and outputs Q-values for all possible actions (in this case 5 actions).
The input arguments to the model are described as below. All inputs have to be normalized between -1 and 1.
The output of the model is a vector 5 elements which indicates Q-values for all possible actions. the best action
is an action with the highest Q-value.
o = (value for discharge, value for half discharge, value for idle, value for half charge, value for charge)
1global_model = nn.Sequential(nn.Linear(6,256),
2 nn.ReLU(),
3 nn.Linear(256,128),
4 nn.ReLU(),
5 nn.Linear(128,5))
6global_model.load_state_dict(torch.load('global_model_q_network.pt'))
The algorithm was developed by
AI4E research Group at UGhent. If you need any support,
don’t hesitate to reach out to us at
seyedsoroush.karimimadahi@ugent.be1@article{madahi2024multi,
2 title={Multi-source Transfer Learning in Reinforcement Learning-based Home Battery Controller},
3 author={Seyed Soroush Karimi Madahi, Toon Van Puyvelde, Gargya Gokhale, Bert Claessens, and
4 Chris Develder},
5 booktitle={Proceedings of the 11th ACM International Conference on Systems for Energy-Efficient
6 Buildings, Cities, and Transportation},
7 year={2024}
8 }
