Improving Solar and PV Power Prediction with Ensemble Methods

TL;DR

This paper improves photovoltaic power prediction accuracy by combining time series models and weather forecasts using ensemble methods, and demonstrates implementation on a Raspberry Pi for microgrid energy management.

Contribution

It introduces a novel ensemble framework integrating various models and weather data for short- and medium-term PV power prediction, implemented on low-cost hardware.

Findings

Enhanced prediction accuracy with ensemble methods.

Successful deployment on Raspberry Pi 3.

Improved microgrid control performance.

Abstract

Estimation of the generated power of renewable energy resources is in general important for planning operations as well as demand balance and power quality. This paper addresses the problem of the estimation of the short-term (3-hour ahead) and medium-term (1-day ahead) generated power of a photovoltaic plant. Firstly, the design of day-ahead solar radiation predictors is investigated with different setups of time series models, and with their combinations with the weather forecast services using ensemble methods. Support Vector Machine methods are also adopted in this stage, to cluster data. Secondly, under a similar ensemble framework, the generated power prediction is investigated. The whole generated power and solar radiation prediction tasks are then implemented on a low-cost, embedded mini PC module Raspberry Pi 3. As an application, the prediction is employed in the control system of a typical microgrid settings focusing on energy management problem. The impact of the quality of generated power prediction on the performance of the controller is also evaluated in this paper.