Solar Power Prediction Using Iterative Network Pruning Technique for Renewable-based Microgrid

This paper proposes a simple and efficient day-ahead solar power prediction method by improving our previous feedforward neural networks (FNNs) based approach, which is applied to a renewable-base microgrid (MG). The proposed prediction algorithm uses only public available weather data such as day-ahead solar irradiation. Based on weather clustering such as sunny/cloudy/rainy, the multiple FNNs are developed for different weather patterns and time zones. An iterative pruning (IP) algorithm is introduced in the previous FNNs structure to achieve efficient computation. Confidence intervals (CIs) indicating the reliability of the forecasting are displayed in the forecasting result. The values of the CIs are effectively provided for the local energy management and robust power system security in MG. The computation times are improved about 90% against the previously proposed method. At the same time, the prediction accuracy is maintained against previous methods. The rate at which the measured value appears outside the confidence interval is 10% or less.