Day Ahead Pricing Forecasts for Short Term Scheduling in Power Markets — A Deep Learning based approach
--
Written by Akshit Gupta and Cremers Sho
In this blog, we will go over deep learning-based RNNs (specifically LSTMs) to forecast day-ahead electricity prices in the context of power markets. The work is mainly based on the approach highlighted in [1] and uses publicly available real-world datasets for weather and electricity prices for training and evaluation. Our results show that RNNs (bidirectional LSTMs) are a powerful tool for forecasting electrical prices with quantifiable uncertainties. In doing so, we were successfully able to replicate the results of [1], albeit on a different dataset.
All our code is open source and available on Github here.
Motivation (The who, what and why?)
In the modern world, electricity is a tradable quantity. The central authority regulates power grids in most countries to ensure transparency in the power markets. Since electricity storage is still costly, a balance between supply (generation at source) and demand (consumption at the sink) is desired in power grids. Due to these constraints, electricity generators and consumers (hereon, called energy actors as per [1]) submit pricing bids to the central authority at a fixed particular time of the day based on their multi-step ahead probabilistic forecasts to maximize their return on investment. These electricity prices are dependent on a large number of extrinsic factors such as renewable energy generation, weather conditions, and the season of the year.
For those interested, an in-depth explanation of these concepts is given here.
So, this work caters to the energy actors (“who”) as described in the preceding paragraph who want accurate electricity pricing (“what) in power grids a day in advance from a model in order to maximize their RoI (“why”).
Theory ( The How? P1)
Being almost a trillion-dollar industry [2], introduced only a few decades ago, naturally, a few million smart people in the world have come up with various ways to get the most accurate pricing forecasts. Some of the earlier and current works in this domain, such as ARMA (Autoregressive…
