An Ensemble Framework For Day-Ahead Forecast of PV Output Power in Smart Grids

Muhammad Qamar Raza,N. Mithulananthan,Jiaming Li,Kwang Y. Lee,Hoay Beng Gooi

DOI: https://doi.org/10.48550/arXiv.1811.09337

2018-11-23

Abstract:The uncertainty associated with solar photo-voltaic (PV) power output is a big challenge to design, manage and implement effective demand response and management strategies. Therefore, an accurate PV power output forecast is an utmost importance to allow seamless integration and a higher level of penetration. In this research, a neural network ensemble (NNE) scheme is proposed, which is based on particle swarm optimization (PSO) trained feedforward neural network (FNN). Five different FFN structures with varying network complexities are used to achieve the diverse and accurate forecast results. These results are combined using trim aggregation after removing the upper and lower forecast error extremes. Correlated variables namely wavelet transformed historical power output of PV, solar irradiance, wind speed, temperature and humidity are applied as inputs to the multivariate NNE. Clearness index is used to classify days into clear, cloudy and partial cloudy days. Test case studies are designed to predict the solar output for these days selected from all seasons. The performance of the proposed framework is analyzed by applying training data set of different resolution, length and quality from seven solar PV sites of the University of Queensland, Australia. The forecast results demonstrate that the proposed framework improves the forecast accuracy significantly in comparison with individual and benchmark models.

Signal Processing

What problem does this paper attempt to address?

The problem that this paper attempts to solve is that the uncertainty of solar photovoltaic (PV) power generation output poses a significant challenge to the design, management, and implementation of effective demand - response and management strategies. Therefore, accurate PV power generation output prediction is crucial for achieving seamless integration and a higher level of penetration. This paper proposes a neural network ensemble (NNE) scheme based on a feed - forward neural network (FNN) trained by particle swarm optimization (PSO), aiming to improve the accuracy of PV power generation output prediction.

Specifically, the paper addresses this problem through the following points:
1. **Proposing a new prediction framework**: Five FNN structures with different complexities are used to achieve diverse and accurate prediction results, and these results are combined after removing the extreme values of upper and lower prediction errors through the pruning - aggregation technique.
2. **Selection of input variables**: Relevant variables such as the wavelet transform of historical PV power output, solar radiation, wind speed, temperature, and humidity are applied as inputs to the multivariate NNE.
3. **Classification method**: The clearness index is used to classify days into sunny, cloudy, and partly cloudy days to analyze the performance of the NNE in each season.
4. **Test cases**: Test cases are designed to predict the solar output of these days selected from all seasons.
5. **Performance analysis**: By applying training data sets with different resolutions, lengths, and qualities, data are obtained from seven solar PV sites at the University of Queensland in Australia to analyze the performance of the proposed framework.
6. **Comparison with benchmark models**: The prediction results show that, compared with single models and benchmark models, the proposed framework significantly improves the prediction accuracy.

In summary, the core objective of the paper is to improve the accuracy of solar photovoltaic power generation output prediction by proposing an improved prediction model, thereby addressing the challenges brought by the uncertainty of solar power generation output.