Aims

In this talk, I would like to

• Introduce some key challenges facing the energy industry, and explain why Data Science and Python could be the key to solving these
• present some interesting approaches and results from our work on energy demand forecasting
• most importantly, show how deep learning models can be built to systematically incorporate a range of often disjoint time series modelling concepts

Background

Increased adoption of domestic battery storage, electric vehicles and solar by the households, as well as increasing reliance on distributed wind and solar generation, are making both sides of the supply-demand equation increasingly dynamic. At the same time, adoption of smart meters has given us access to vast amounts of previously unavailable data.

The challenge

At Octopus Energy, we are developing new technologies in the energy space. Two key problems I will address in this talk are:

• predicting demand for hundreds of thousands of individual consumers with varying availability of historic data
• modelling the probabilistic nature of both energy demand and energy prices to quantify risk to consumers and suppliers

TIme Series Forecasting and Deep Learning

The field of time series forecasting is vast and often confusing. From "classic" AR models to recent developments such as LSTMs, it can be difficult to know where to start. We leverage the power of pandas, scikit, keras, tensorflow to build models of increasing complexity and power. Through this talk, I would like to explore some core concepts and show how approaches from different fields can be unified successfully. In particular, I would like to show how:

• a forecasting problem can (and should) be seen as a regression problem
• embeddings provide a powerful way to implement transfer learning, and are the deep learning equivalent of latent variable models
• (auto-)encoders can take care of auto-regressive features
• we can turn any deep learning model into a probabilistic one, thus bridging the gap to a number of stochastic modelling approaches
• we can combine concepts from machine learning and finance to better model correlations and tail-dependencies between energy demand and energy price