Deep Latent Generative Models for Energy Disaggregation
Authors: Gissella Bejarano, David DeFazio, Arti Ramesh850-857
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this work, we propose a deep latent generative model based on variational recurrent neural networks (VRNNs) for energy disaggregation. Our model jointly disaggregates the aggregated energy signal into individual appliance signals, achieving superior performance when compared to the state-of-the-art models for energy disaggregation, yielding a 29% and 41% performance improvement on two energy datasets, respectively |
| Researcher Affiliation | Academia | Gissella Bejarano SUNY Binghamton gbejara1@binghamton.edu David De Fazio SUNY Binghamton ddefazi1@binghamton.edu Arti Ramesh SUNY Binghamton artir@binghamton.edu |
| Pseudocode | No | The paper includes architectural diagrams and mathematical equations, but does not present any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | Yes | 1https://bitbucket.org/gissemari/disaggregation-vrnn |
| Open Datasets | Yes | We evaluate our model on two real-world energy datasets: i) Pecan Street Inc. Dataset (DATAPORT) (dat 2016), and ii) Reference Energy Disaggregation Dataset (REDD) (Kolter and Johnson 2011). |
| Dataset Splits | Yes | We split the total number of instances into training, testing, and validation sets in the ratio 50%:25%:25%, respectively. We record the performance metrics in the validation set every ten epochs to detect and prevent overfitting. |
| Hardware Specification | No | The paper does not specify the exact hardware used for experiments, such as CPU or GPU models, or memory specifications. |
| Software Dependencies | No | We develop our model on the original VRNN implementation (Chung et al. 2015) in Theano. However, specific version numbers for Theano or other software dependencies are not provided. |
| Experiment Setup | Yes | The weight matrices of all layers are randomly initialized using a uniform distribution. The LSTM-cell diagonal matrix that captures the interaction between the recurrent states ht 1 and ht is initialized randomly from a normal distribution ensuring its orthogonality. The initial hidden state of the recurrent neural network is initialized to 0. We experiment with different activation functions for θµ and find that Re LU activation function for θµ works better for some buildings while for others the linear activation function works better. For θσ and coef 1, we apply a softmax and softplus activation functions, respectively. ... We use 5-30 mini-batches. We report the average scores from three different train-test-validation splits across both datasets. |