Interval forecasting of solar power generation

Abstract

Solar energy is one of the most promising renewable energy sources for electricity generation due to its long lifespan and low maintenance fee. The implementation of solar energy generation capable of reducing the carbon emission from energy generation sector. Moreover, global solar generation capacity has increased by 24% from 2018 to 2019 and the capacity will definitely increase when solar panel with higher efficiency and lower carbon footprint being invented. However, large scale of intermittent and variable solar power generation will considerably impact to the power grid system. In general, solar forecasting method can be divided into physical method (Numerical Weather Prediction (NWP) based), statistical method (Machine Learning) and hybrid method. Physical approach will always have its major drawback when forecasting short-term PV power due to time consuming on NWP stimulation. Statistical forecast based on historical data are popular recently due to its simplicity and adaptability. Thus , in this project, statistical method, especially machine learning-based PV forecasting methods will be the main focus. Moreover, traditional deterministic forecasting method only provide expectation value of PV power, which can lead to inevitable error and deteriorate the reliability of future PV power prediction-based energy management system. In this project, original point forecasting method such as deep neural network (DNN), ensemble DNN and long short-term memory (LSTM) are extended to probabilistic forecasting to achieve uncertainty measurement on the forecasted results. Gaussian Process (GP) will also be studied in this project for its similar characteristic. Numerical experiments are carried out based on irradiance data collected in Singapore and power data collected in Australia. Method such as Prediction Interval Coverage Probability (PICP), Prediction Interval Nominal Confidence (PINC) and Interval Score (IS) will be used to evaluate the results from different machine learning model. Results show LSTM model out-perform DNN model and GP model in both point and probabilistic forecasting on large datasets. GP results can still be taken account since it required less data compared to other 2 model