Error Spectrum Analysis of Solar Power Prediction for Deakin Microgrid Digital Twin

In 2021, the AUD 23 million Deakin University Renewable Energy Microgrid commenced operation. To date, it has the largest solar farm at an Australian University- one that is capable of producing 7.25-megawatt of power. In the year following that, a web-based digital twin was developed with the aims of providing operators with intelligence and insights through AI-driven capabilities. A key component of these AI-driven capabilities is a fast algorithm that predicts power generation from weather data for rapid decision-making and frequent health assessment. In particular, the accuracy of the prediction algorithm needs to be high and consistent, regardless of the magnitude of the power generated. To this end, we tested a number of commonly-used ML algorithms. From our experiments, Extreme Gradient Boost (XGBoost) and Random Forest (RF) are capable of training 7500 samples well within a few seconds with reasonable accuracy. With this observation, we further tuned the two methods and compared them for accuracy and consistency using data from different sections of the solar farm, mainly the three main blocks (each with over 6000 solar panels) and the four research arrays (each with 144 solar panels). Given a similar (or much less) computation time, RF outperformed XGBoost in terms of prediction accuracy using the new accuracy profile metrics we have introduced: namely the x-percentile Closeness Scores and the x-percentile Absolute Errors. These new metrics measure the performance by reporting the whole error spectrum instead of just a single value.