Advanced Wind Power Predictions Based on Deep Learning

Abstract

The growing problem of air pollution and the shortage of fossil fuels call for more research on renewable energy. Wind energy, one of the cleanest renewable energy sources, has contributed to an increasing share of electricity generation in many countries since recent decades. However, the stochastic nature of wind makes the wind energy production highly variable, posing a great challenge to grid management and limiting the penetration of wind energy in power grids. Providing reliable and accurate wind power predictions plays a critical role in reducing the waste of wind energy.

Today, the vast amount of operational data collected through sensors installed on wind turbines provide an unprecedented opportunity to apply advanced data mining techniques to wind power predictions. In such data-driven wind power prediction approach, particularly, there are four issues that need to be investigated and properly addressed. First, the harsh operating environment of wind turbines can cause occasional sensor failures, which in turn may result in data loss and seriously affect the usability and performance of data-driven prediction models. The missing data problem should be first handled in data preprocessing. Secondly, based on huge amounts of industrial data, training data-driven models takes a long training time. High training efficiency needs to be ensured. Thirdly, as the core concern in a prediction problem, the prediction model should demonstrate high accuracy. Finally, since no prediction model can always be perfectly accurate, the uncertainty of prediction results should be estimated reliably and with confidence.

To address the aforementioned four aspects, this thesis provides a comprehensive and in-depth study of the data-driven modeling approach for wind power prediction by proposing advanced deep learning based methods. In the missing data problem, a two-stage deep learning-based missing data imputation framework is first proposed to accurately recover incomplete data that present difficult missingness patterns. To improve the modeling efficiency of deep learning-based prediction models, a transfer learning framework is proposed that allows better utilization and organization of training data from multiple wind turbines, as well as rapid determination of hyperparameters for deep neural networks. To enhance the prediction accuracy, a novel convolutional recurrent deep neural network combined with a two-layer clustering process is introduced. Finally, a deep learning framework that produces and aggregates point predictions and multi-interval length predictions is proposed to achieve reliable and confident probabilistic wind power predictions.

The effectiveness of proposed methods in the thesis has been validated with real industrial datasets collected from several wind farms.

Date of Award	19 Jul 2021
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Zijun ZHANG (Supervisor)