A comprehensive framework for effective long-short term solar yield forecasting

Abstract

Due to the variability of Photovoltaic (PV) output, a forecasting framework is essential for grid connected PV plants to ensure a stable and uninterrupted power supply. Among existing prediction and forecasting algorithms, only some have attempted to provide a holistic framework for short and long-term forecasting of PV yield together using automated input feature selections and data cleaning features. Furthermore, it has been identified that many existing algorithms only predicted PV output instead of forecasting in future times; therefore, their reported accuracy needs to be upheld in forecasting scenarios. This paper has proposed a framework to streamline solar yield forecasting for both the short and long term to ensure effective integration of PV plant output with the main grid. The proposed framework has used a novel combination of XGBoost (eXtreme Gradient Boosting), time series seasonal decomposition and rolling LSTM (Long- and Short-Term Memory) model to address the need for a comprehensive forecasting framework in hourly, daily and yearly periods. Based on our experiment result, the developed framework has performed in 98% − 95% prediction accuracy with less than 0.15% normalized Root Mean Squire error (nRMSE). The framework has performed in 89%- 87% forecasting accuracy with less than 0.45% nRMSE. Both the prediction and forecasting performance of the proposed model have outperformed many benchmarks forecasting frameworks, including Long short-term memory (LSTM) based recurrent neural network (RNN), Full RNN (FRNN), Neural Network Ensemble (NNE), Neural Network with AdaBoost, and many more as detailed in our comparative study section.

Keywords

• Prediction
• Forecasting
• Photovoltaic
• Machine Learning
• Framework