Renewable energies, such as wind and solar, are critical pillars in the pursuit of global carbon neutrality, and the installation capacity has steadily increased in recent decades. This upward trend provides a solid foundation for powering our society with clean and renewable energy, while also mitigating the environmental pollution caused by fossil fuels. However, volatility of renewable power poses new challenges to power grid operations, particularly with regard to safety and stability. Therefore, renewable power prediction has become a central topic, as its results can help grid and power system operators better manage this uncertainty.

Recently, widespread deployment and continuous advancement of the data acquisition systems in renewable power plants offer an unprecedented opportunity for studying deep learning-based renewable power prediction techniques. Apart from the advanced model structure innovations, which is extensively studied in recent studies, data-efficient frameworks of utilizing the information in renewable power plant data should also be addressed for a better prediction performance. To fill up this important research gap in the renewable power prediction field, four research works are conducted in this thesis, which are presented as follows.

Firstly, for better solving the short term renewable power prediction problem, we introduce a new data usage paradigm, using data of both high and low sampling resolutions as model inputs. To organically incorporate data of two sampling resolutions as inputs and tackle the extreme high dimension of features induced by using high sampling resolution data in renewable power prediction modeling, we propose a novel bilateral branch learning based modeling framework, which includes two data feature engineering branches and one prediction module. Based on results of a comprehensive computational study, we verify that our proposed framework achieves the state-of-the art performance as it beats a large set of classical data-driven and recent deep learning-based methods considered in this study.

Secondly, to consider data of multiple sources from different renewable power plants and extract spatial-temporal features for a better prediction performance, a two-channel deep network modeling method for renewable power predictions is developed via leveraging both the renewable power plant data and plant geo-information. To advance predictions with such input, a deep graph attention convolutional recurrent method, which develops one novel deep network channel for engineering high-level latent features from high-dimensional inputs and another classical feature selection channel for directly engaging valuable attributes, is proposed. Comprehensive computational experiments are conducted to verify the value of such modeling development by comparing it with a set of competitive benchmarking models. The results show that the proposed method achieves the state-of-the-art prediction performance.

Thirdly, to preserve the privacy while using the data of multiple sources, a novel bi-party engaged data-driven modeling framework is developed to enable efficiently learning local and global latent features serving as decentralized data for data-driven modeling with privacy-preserving. The proposed framework enables capturing spatial-temporal patterns among multiple sites to further benefit modeling in renewable power prediction tasks. Meanwhile, the framework preserves the data privacy via isolating local data in local clients. Next, a comprehensive computational study is conducted to benchmark it against famous baselines. Results show that the proposed framework achieve at least 3% improvements on average.

Finally, the privacy preserving learning framework is extended to enable the probabilistic predictions. To perform such a task, an advanced federated deep mixture density network is developed. The proposed method consists of multiple clients and one server. Each client transforms local data into local latent features using a deep learning-based local feature extractor and provides probabilistic forecasts using a local forecasting model. The server is designed for two purposes: 1) aggregating the knowledge from clients in the models, which are then dispatched to each client; 2) helping the local feature extractors identify domain-invariant features via a discriminator. A preliminary exploration of the theoretical property of the proposed framework is conducted to partially explain its advantages. Additionally, computational studies are performed to further confirm the advantages of the proposed method as it offers best performances in most of application scenarios.