I would like to start a discussion. I am dealing with a consumption forecasting project and analyzing a database composed of more than 60k id counters.
The goal is to create a system able to model the series efficiency and avoiding high time-consuming. I kept the following two approaches:
- Sarima model applied to each time series, but it takes too much time (1.2 months) because of the Grid Search Approach useful to fit all the parameters in each id.
- Fastai library, able to create a Tabular Data Bunch and a unique model for all the series, which has lower performance than the classical approach but it is advantageous by the time-consuming side (only 4 hours training).
In the first approach, I only used the historical consumption values, while, for the second approach I considered more features such as temperature, humidity, building id, apartment id, city, and so on. Actually I am trying to create a model using Keras and Tensorflow, in order to control and create the model, but I don't see improvements.
It seems I am not able to reach a good settlement between performance accuracy and fast model creation. Do you have any suggestions? Any similar experience on this?