Load Forecasting with Deep Learning

Challenge

Forecasting daily electricity consumption is a central challenge for many energy companies, as it directly impacts the management and provision of the electricity to be purchased on the market. Deviating forecasts can result in electricity needing to be bought at higher prices or sold back at below-market prices. Load forecasting is often handled through black-box products, which can lead to unfavorable side effects like vendor lock-ins and a lack of transparency. Our client, an international energy company, faced the challenge of benchmarking their existing black-box solution against open-source tools.

Approach

Load forecasting is often conducted with simple neural networks, which are well-suited for predicting the periodic components of daily electricity consumption while considering other influencing factors such as holidays, public holidays, and weather parameters. With these advantages in mind, we developed a deep learning model consisting of multiple hidden layers compared to simple neural networks to additionally model high-dimensional interactions between input factors. Furthermore, individual models were repeatedly trained on different data partitions to create an ensemble forecast.

Result

The developed model showed a significant improvement in forecast accuracy compared to the existing black-box solution in 50% of the sales areas within the control zones. Additionally, the variance of the forecasts was significantly reduced through the application of ensemble bagging.