Start-to-end guide on building a load forecasting model in PI

Blog Post created by bbachiega on Nov 22, 2017

An accurate power load forecasting model is fundamental for operations and planning of any Transmission & Distribution company. It proves to be a valuable tool especially during hot summer days and cold winter nights where the price of electricity fluctuates and often multiplies in value. An accurate forecast can save a company from underbidding and having to buy electricity at market prices in these conditions or overbidding, leaving paid electricity to go unused. A load forecast is often available by the ISO and the EMS, but these are closed down systems that don't allow for any changes or improvements. This guide will take it a step further and create a forecasting model at a substation level, something valuable for many reasons that we will cover later.


Forecasting applications can also reach other fields with high potential for optimizations (read $ signs) such as predicting demand growth, future prices, solar and wind power output based on weather conditions. Multivariate analysis and machine learning can provide interesting insight and invaluable information on these big, complex problems.


Let's start with our goal. The image below is the screenshot of a PI Vision display showing actual load (green) and forecasted load (red) for a 7 day period:




This guide will be broken down into different parts as detailed below:


Part I (Building a load forecasting model in PI - Part I ):

Data preparation - The data used to build the model is assumed to be in the PI System and modeled in AF. For the sake of keeping things simple we will only use temperature and time for predicting MWs and the PI Integrator for AZURE will be used to:

  • Cleanse
  • Augment
  • Shape
  • Transmit


Part II (Building a load forecasting model in PI - Part II ):

Forecast model - AZURE Machine Learning will be used in this example although the focus is not this tool but the process of integrating with PI and building a forecast model. Other platforms could be used instead such as Hadoop, SAP HANA, R, etc. The process to build a forecasting model is usually comprised of:

  • load data
  • train
  • score
  • deploy



Required for this example:

  • PI Data Archive 2015+         //Future data support
  • PI Vision                               //Visualization
  • PI Integrator for AZURE       //Data integration
  • MSFT AZURE subscription  //machine learning