PROLOAD

A power load forecasting module for medium voltage electricity distribution networks

In combination with Supervisory Control And Data Acquisition (SCADA) systems, PROLOAD is able to predict the aggregated consumption of electric power at different measurement points.

Configurability is emphasized by virtue of selectable time scales for the forecasting window and the underlying prediction engine.

The tool also features self-learning capabilities in the form of retrofitted inputs to the selected prediction algorithm.

The tool suite incorporates additional functions capable to analyze and infer load similarities and patterns within the historical database of load traces under utilization.

Sophisticated fusion and correlation of relevant side data (e.g. weather forecast) along with historical load traces is achieved thanks to the flexibility of the predictors, which further enhances the performance of the tool.

Contacta con nosotros

Introduction

With the proliferation of smart metering devices and the progressively increased complexity of electricity power distribution networks, operators and other relevant stakeholders have been urged to implement and deploy innovative tools for electric load forecasting so as to:

  • Improve the distribution efficiency by means of active power consumption monitoring and preemptive detection of non-technical and commercial losses.
  • Ease and enhance the dimensioning process of the distribution network itself by identifying electric power demands and needs before they actually occur.

Technological approach

The PROLOAD module developed by TECNALIA is based on a wide portfolio of techniques and algorithms springing from Soft Computing and Artificial Intelligence, which allow for the short- (hour, day ahead) and long-term (month, year ahead) forecasting of electricity loads in medium voltage distribution networks. The module can be easily configured to operate with any of the following techniques:

  • Neural processing methods, not only naïve versions of conventional techniques (e.g. Back-Propagation NN, MLP, ANN), but also specialized neural approaches designed ad-hoc to meet specific needs and requirements of the scenario at hand (speed, accuracy).
  • Regressive models such as ARMA, ARMAX, ARIMA, SARIMA and GARCH, the latter for short-term forecasting.
  • Cross-validation based ensembles of decision trees (Random Forest).
  • Support Vector Regression with linear, polynomial, radial basis or sigmoid kernel functions.
  • Time Series Analysis with Self-Organizing Maps (SOM): under development.
  • Other classical techniques such as Kalman filtering, least squares regression and Croston’s method for extremely intermittent load patterns.
  • Ensembles of prediction engines with iterated cross-validation training.