Short-Term Electric Load Forecasting with LSTM and Gradient Boosting

A grid operator has to forecast the electricity demand of the next one to two days every day to schedule generation and buy power. This project forecasts the hourly national electric load of Germany for the next 48 hours, issued every day at 00:00 UTC, from two weeks of past load, the calendar (hour, weekday, public holidays, bridge days) and the temperature. Two machine-learning models - LightGBM gradient-boosted trees and a small LSTM encoder-decoder that runs on a laptop CPU - and their ensemble are compared with the classical forecasters used by utilities: seasonal naive and multiple linear regression (MLR, including Tao Hong's vanilla benchmark). The models are trained on 2015-2018, chosen on 2019 and tested once on January-September 2020, the year of the COVID-19 lockdown. An 80 % prediction interval, a web application with five databases, a GNU Octave cross-check (run) and a complete MATLAB version (written, not run) are included.

Category: Electrical and Electronics Engineering (Power-System Operation, Energy Management, Load Forecasting, Machine Learning, Deep Learning, MATLAB, GNU Octave, Python)

Electrical & Electronics Engineering Project EE01 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.