Forecasting energy demand

The Case:

On May 8th, 2019 India’s power demand soared to a record high around four months before consumption is usually the highest in a year. Although the energy produced that day was 177.7 GW (12% higher than a year ago) it still did not manage to meet the increased demand of 178.25GW. This spike was attributed to the unusually hot weather.

Such drastic spikes in energy demand can cause energy shortages (which need to be addressed by importing more expensive energy from abroad) and even blackouts. This can have a negative impact on the entire economy.

The Challenge:

Predicting energy demand is no easy task either. The demand for energy is influenced by a wide range of factors such as the weather, socio-economic factors such the growth in GDP, holidays and many more.

Traditional statistical models often times fail to accurately predict the energy demand due to this large amount of variables

Techniques such as ARIMA and other methods for time series analysis also often times fail due to the highly seasonal nature of energy demand and the presents of short, mid and long term trends.

The Solution:

To help energy traders better prepare for such events we’ve developed an algorithm for energy demand forecasting based on state of the art neural networks. The algorithm was built with historical demand and weather data provided by an energy trading company. The algorithm uses neural networks to combine data from multiple sources such as the aforementioned weather data and produces an hourly prediction for the next 24 hours.

The Outcome:

The algorithm was tested with a select group of customers of the energy trader and managed to reduce the error of their predictions from 23.19% to 18.16%. This reduction in the forecasting error was predicted to save the energy trader and its clients 150,000 euros per year.