With the rise of smart energy meters and cloud computing, getting detailed insights into your energy usage can be just a few clicks away. In this talk we explain how a deep learning model helped us estimate the actual solar production from redelivered energy for Eneco customers, and explain the technical challenges in getting there.
Smart meters only measure redelivered energy, but can’t measure how much solar energy was used by the household itselff before that. In this talk we will go in-depth into the process of estimating solar production from the energy that is redelivered to the grid at the smart meter, using PyTorch and Horovod. We will go through the entire process from ideation to production-ready: validating the feasibility of the model, gathering a dataset, exploring different solutions and finally training and validating the model.
Next to that there are some technical challenges: To deal with the scale of data we used the distributed deep learning framework Horovod for training, and Spark for distributed predictions. We show how we use MlFlow for keeping track of our model iterations and finally, we will walk you through the data-engineering required to get the data presented in our mobile app!
Come join us to learn the ins and outs of getting a deep learning usecase to production at scale!