Data-driven Modeling, Controlling, and Forecasting for Photovoltaics Systems

Abstract

The increasing penetration of photovoltaics (PV) power draws a high research attention on PV associated problems to develop a more economical, efficient, and reliable PV system. In this thesis, the following problems are investigated based on data-driven approaches.

·  Maximum power point tracking (MPPT): An MPPT algorithm addresses the problem of finding the maximum power point in PV electrical characteristics curve, thereby maximizing the power transfer from PV panels. Classical MPPT algorithms including P&O (perturb and observation) and INC (incremental conductance) cannot handle PV systems exposed to partial shading conditions where there will be multiple local peaks in the power-voltage (P-V) curve. Hence, an MPPT algorithm capable of handling the partial shading conditions is of great value.

·  Modeling of PV power generation: A PV performance model predicts the outdoor power generation depending on environmental factors including plane-of-array (POA) irradiation, PV cell temperature, solar angle-of-incidence (AOI), and solar spectrum. An accurate and easy to implement model is essential for the planning, monitoring, and diagnosis of PV systems.

·  Evaluation of novel PV cell technology: PV cells made from upgraded metallurgical grade silicon (UMG-Si) are promising alternatives to high-cost conventional solar grade silicon. However, the advantage of low cost of UMG-Si will be impaired if outdoor efficiency is reduced by degradation of quality. Thus, more reporting of field performance benefits further development of UMG-Si PV systems.

·   Short-term forecasting of solar irradiation: The uncertainty of solar power productions raises more challenges for the operation and planning of power systems including PV systems. The PV power production highly depends on incident solar irradiation; thereby, accurate forecasting of solar irradiation is beneficial to the operation of PV power plants. 

The main contributions of the thesis are outlined as follows:

1)  A novel model-free solution algorithm, the natural cubic spline guided Jaya algorithm (S-Jaya), is developed to efficiently address the MPPT problem of PV systems exposed to partial shading conditions. Jaya is a variant of swarm intelligence, and it is free of algorithm-specific parameters. In the S-Jaya algorithm, a natural cubic spline prediction model is integrated into the iterative update of candidate solutions (operating voltage settings). Such an extension can reduce the negative updates thereby improving the MPPT performance. Results of simulation studies and experiments considering a variety of partial shading conditions show that the S-Jaya algorithm provides a faster convergence speed, smaller oscillations, and higher overall tracking efficiency against benchmarking models including the generic Jaya algorithm and the particle swarm optimization (PSO) algorithm.

2)   A method for improving the accuracy of artificial neural network (ANN)–based prediction of PV power generation is developed. In addition to the generally utilized information of POA irradiation and module temperature, solar zenith angle and azimuth angle are introduced as new inputs to account for impacts of solar AOI and solar spectrum. Numerical studies based on data from two different PV sites illustrate the improvement in accuracy of the proposed ANN model against benchmarking models. It is also demonstrated that solar zenith angle and azimuth angle are useful for support vector machine (SVM) regression and Gaussian process regression (GPR) to model PV power generation.

3)  The methodology for comprehensively evaluating the long-term field PV performance is presented and applied to a 1.26 kW grid-connected UMG-Si PV system. A comparison of performance ratio and degradation rate to conventional crystalline silicon-based PV systems suggests that the performance of the UMG-Si PV system is comparable to conventional systems. This reporting provides a preliminary understanding of field performance of UMG-Si PV system and offers hands-on experience for better and wider application of UMG-Si PV modules. This study also systematizes and rationalizes the necessity of a data selection and filtering process to improve the accuracy of degradation rate estimation.

4)    A data-driven framework is proposed for the short-term forecasting of solar irradiation at a targeted site considering its time-series and those observed at neighboring sites. Famous data-driven approaches including the boosted regression trees (BRT), ANN, SVM, and least absolute shrinkage and selection operator (LASSO) are applied to model the spatial dependence among the solar irradiation time-series and to forecast the solar irradiation at a targeted site. Inputs selection is conducted based on the relative importance estimated by the BRT model to reduce the model dimensions. A comprehensive comparison of data-driven forecasting models is performed and the BRT model is considered as the best candidate. Moreover, computational results of multi-steps ahead forecasting demonstrate that the BRT model also outperforms the widely considered benchmarking models including the persistence model, autoregressive (AR) model, and autoregressive exogenous (ARX) model in terms of forecasting accuracy.

This thesis contributes to develop a novel MPPT algorithm for PV systems exposed to partial shading conditions, improve the accuracy of ANN-based modeling of PV power generation, investigate the long-term field performance of a low-cost UMG-Si PV system, and propose a data-driven framework for short-term solar irradiation forecasting. These contributions are beneficial to the development of a more economical, efficient, and reliable PV system.