solution for forecasting electricity demand

Rapid climate change has resulted in record-breaking hot summers, while extreme cold temperatures, on the other hand, have resulted in colder winters. Consequently, the number of buildings that use air-conditioners and heating appliances is increasing every year, HVAC are used more frequently and their combined energy usage has an increasingly negative impact on nationwide energy distribution. These energy-intensive devices are responsible for a large portion of peak load demand and this demand continues to set new records every year.

A way to fight peak load demand is to plan for securing adequate generation, transmission, and distribution capacities, while accurate peak forecasts improve decision making and improve reliability of the energy system. Accurate peak forecasts not only improve decision making, reliability and energy efficiency of the power grid, but also prevent the waste of energy and with helps environmental issues by establishing plans for the use of renewable energy and enables companies to store the renewable energy when it is available and use it during peak demand. Future peak load is not deterministic and it depends on several uncertain factors including weather conditions, economic activity, population, etc.

 Peak-load forecasting is largely divided into two approaches:

  • Time-series approach: using system load information over time. Most time-series models assume that time-series data are stationary and that the trends don’t change much over the time.
  • AI-based approach: using machine learning and deep learning, which creates a predictive model on the basis of information about various factors, such as weather information (temperature, wind speed, sunshine, etc.), day of the year, economic activity, etc. The predictive model then finds statistically relevant correlations among different factors that affect peak demand and is thus able to predict the time and the size/shape of the peak. In addition, by comparing predicted results with actual measured values, the accuracy of the predictive model can be continuously improved.

The case study, on the right side, describes, how we helped a client Stem Inc., the leader in the field of AI based storage-as-a-service, build an AI based predictive model for peak load forecasting in Ontario, Canada. Stem uses large battery storage to store renewable energy when available (such as from rooftop solar, wind farms, etc.). During peak demand, they can they disconnect an entire building or production facility from the grid and power it off the batteries, thus bringing significant savings in peak demand fees reduction and improving the overall efficiency of the electric grid.

The batteries are expensive and a significant financial investment and it is crucial they are used optimally. It can take several hours to charge the battery and if the battery is discharged on a “false” peak, there will be no more charge left for the real daily peak load, meaning we are not managing this expensive resource optimally. The goal of our AI peak load prediction solution was to hit daily peaks forecasts with high precision to enable the batteries top be discharged during the highest peak demand of the day.

A Case Study at a Glance


  • Fluctuations in electricity demand and supply
  • Daily Peak prediction
  • Integration of large machine learning dataset


  • Reducing electricity cost for the end users
  • Improving client’s profit margin
  • A better utilization of renewable energy sources, improving efficiency


  • Python with different frameworks (Django, Falcon, Flask, etc.)
  • Java for High-performance Microservices
  • AWS Infrastructure (DynamoDB, etc.)