This report outlines lessons that Meridian Energy Australia and the University of Melbourne have recently learnt regarding clustered wind turbine-based forecasting and wind turbine cross correlations for the Mt. Mercer wind farm.
Objective: Demonstrate the ability to submit five-minute ahead self-forecasts via AEMO’s web based MP5F API.
Details: Our a priori hypothesis for this project was that sub-diving the wind farm into groups of similar producing wind turbines and summing a forecast for each group would yield an overall more accurate forecast than a single power forecast. This hypothesis was based on the assumption that our errors, provided they were “truly random”, would cancel each other out when aggregating and our forecast accuracy would increase.
The trade-off for this gain in accuracy, would be the increased difficulty in forecasting (namely requiring 64 data input streams for each turbine versus the single power data input stream). However, in an offline forecasting scenario (where live data availability was not an issue), our forecast error decreased with an increasing number of clusters only to a point, as shown in Figure 1. Here our “Ensemble” forecast in red is a combination of lightGBM and xgboost, whereas our lightGBM forecast in yellow is simply the lightGBM forecast. These two machine learning approaches were selected for two reasons, namely they were robust (able to handle missing values) and they performed well in preliminary tests. After 18 clusters, the forecast accuracy did not further increase. Note that our Figure only shows to 20 clusters for readability, but this trend continued to the point where our analysis was terminated.