Closed bailingbird closed 2 years ago
The code for Monte Carlo dropout is not included in this repository. To implement it: 1) add dropout layers to the model 2) train the model with a specific dropout rate (e.g., 0.05 or 0.1) 3) at test time, keep the dropout layers on and feedforward multiple times (e.g., 100 times) and calculate the variance for each hour in the horizon (term1 in eq. 8). 4) calculate term2 in eq. 8 on a validation set for each hour in the horizon. A coefficient may be needed to ensure proper empirical coverage. 5) combine term1 and term2 to obtain the total variance.
Currently I would suggest implementing it in PyTorch, which is much easier.
Fig.8 shows 95% prediction intervals for a winter week. Then, how to modify the code in order to achieve the Probabilistic Forecasting? Thanks very much.