Forecasting Solar Power Time Series: Strategies For Multi-Modal Data Fusion, Feature Relevance, and Sparse Data Management

The forecasting of solar photovoltaic power (SPVP) is a significant challenge. Solar is the least reliable renewable energy, as it depends on the weather, among other things. However, it is also one of the cheapest sources if it can be harnessed, particularly during daylight hours when people work and use electricity. The ultimate aim of forecasting solar power using deep learning (DL) techniques is to enable the aggregated use of solar power stations by day, supplemented by alternative sources of electricity whenever solar energy is forecast to fall below a particular level. The more accurate the solar power predictions, the better the use and supply of this valuable resource. This thesis proposes an SPVP forecasting method that applies DL methodologies using real data from multiple solar power stations. SPVP time series data is complex and characterized by variable, dynamic, and multi-dimensional attributes. Consequently the research in this thesis has to address various challenges, predominantly stemming from the inherent characteristics of SPVP data. The multifaceted nature of these challenges includes data variability and non-stationarity, where the influence of diverse environmental conditions, seasonal variations, and geographical factors introduces significant fluctuation and unpredictability into the data. To address this variability, forecasting models that have the capability to adapt and predict based on changing patterns are needed. Additionally, the multi-dimensional nature of the inputs required for precise forecasting poses another hurdle. Accurate SPVP generation forecasting models need to integrate multiple types of data, not only historical generation data but also exogenous variables such as weather conditions. Compounding these challenges is the issue of data availability. Many solar installations, especially new ones or those in less-studied regions, do not have the extensive historical data crucial for training robust forecasting models. Traditional machine learning methods often prove inadequate, as they are limited by their dependence on extensive data manipulation and feature engineering, so the requirements for deep domain expertise—capabilities are not always available. These methods struggle to capture and utilize the dynamic interplay between the factors affecting SPVP generation, and this underscores the need for innovative approaches that can navigate these complexities more effectively. % This thesis addresses these challenges by developing and applying advanced DL models to enhance the accuracy and efficiency of SPVP generation forecasting. Motivated by the limitations of existing forecasting approaches, this research explores innovative DL techniques capable of handling the complexities of SPVP data. To address the challenges posed by data variability, we introduce an aggregated SPVP model with a Wavelet-based-coefficient (Wcoeff) approach that is used for univariate data decomposition to denoise the data. The Wcoeff model redefines the wavelet transform (WT) application to streamline feature extraction. This approach provides a scalable and accurate forecasting solution by mitigating computational complexity yet retaining temporal relationships. Exogenous data is then integrated to enhance forecasting accuracy, and the research addresses the multi-dimensional nature of these inputs through the innovations of the Multilevel Data Fusion and Neural Basis Expansion Analysis (MF-NBEA) model. This model represents a pivotal advance in using DL for SPVP forecasting. Indeed, understanding the most important lagged variables influencing the generation is crucial for refining forecasting models. Given the high dimensionality and evolving nature of the data to be used, a dynamic approach to lagged variable selection and modeling is required. The research develops dynamic feature selection that adjusts to changing conditions and highlights the most predictive variables over time. This adaptability ensures models remain accurate and relevant, even as the underlying data patterns shift. Finally, we introduce a novel methodology that integrates learned knowledge from multiple source domains to address the critical challenges in forecasting accuracy when data is scarce. This innovative transfer learning approach marks a significant departure from traditional single-source forecasting methods. By leveraging the wealth of data available from already established solar power installations, the new methodology enhances the forecasting model's ability to predict solar power output in new locations or locations with limited historical data. The essence of the novelty is in the strategic fusion of knowledge from across multiple domains, utilizing advanced techniques such as average weights fusion and evolutionary optimization based fusion. This thesis makes a significant contribution to the field of DL models and renewable energy forecasting by providing scalable, efficient, and adaptable models. The findings underscore the potential for advanced DL techniques to navigate the complexities of SPVP time series data and offer insights that will facilitate the broader integration of solar energy into the power grid. This work opens avenues for future research to enhance model interpretability, explore cross-domain applications of transfer learning, and further optimize models for real-time forecasting applications.\\