DATA DRIVEN FORECASTING TECHNIQUES FOR ENHANCED PERFORMANCE OF SOLAR PV SYSTEM

Abstract

The global technological advancement has intensified the requirement of additional electrical power sources. This overwhelming demand of electricity redirected the focus of the researchers and utility operators towards the renewable energy resources. The research advancements have facilitated the installations of the power system driven by renewable resources like solar, wind, hydro, etc. Among all the available renewable based technologies, the solar Photovoltaic (PV) systems have achieved the highest adoption trajectory due to their energy independency, long-term investments and cost saving solution. Nevertheless, the solar driven power system persists several challenges related to grid integration, load management and system reliability. The primary cause of these challenges is due to the dynamic atmospheric conditions which are the main drivers of solar PV system. Among various atmospheric parameters, the optimal temperature and solar irradiance are required to maintain the Maximum Power Point (MPP) of solar PV system. Between the two vital atmospheric parameters, the solar irradiance pattern is highly non-linear due to its dynamic variation throughout the day, possessing difficulty in attainment of MPP. Therefore, it is necessary to study the variable pattern of the solar irradiance in order to anticipate the range of solar energy. A similar exhaustive analysis of the electrical load consumption pattern is also required for maintaining the synchronism of supply and demand chain. This highlights the importance of the comprehensive study of the pattern of source and sink of the solar PV system for developing data-driven predictive models. These models will be capable of handling the unexpected fluctuations in both energy generation and consumption stages. The widespread adoption of Machine Learning (ML) techniques across various fields suggest that the integration of the data driven Neural Network (NN) based techniques could be the reliable solution for enhancing the efficacy of the solar PV system. The thesis initially delves into the exploration of effective methodologies for developing robust Maximum Power Point Tracking (MPPT) algorithm, which aims to achieve the maximum power point of the Solar PV system under any adverse atmospheric conditions. The effective MPPT algorithm has been developed by optimizing the internal configuration of Artificial Neural Network (ANN) model. The inherent ability of the ANN models to effectively learn the non-linear pattern of the temperature and solar irradiance dataset has been used to develop the MPPT algorithm. However, the selection of appropriate internal parameters for the training of ANN model is one of the substantial and frequent challenge. To handle this issue, the optimization algorithm v has been integrated with ANN model to obtain its optimal internal parameters. Therefore, a novel ANN-HHO MPPT controller has been developed in which the internal parameters of the ANN model are optimized using recent metaheuristic optimization algorithm i.e., Horse Herd Optimization (HHO). The performance of proposed MPPT controller has been tested and validated under adverse atmospheric conditions including the several complex patterns of Partial Shading Conditions (PSCs). The development of robust data driven ANN based MPPT algorithm requires an adequate solar irradiance data, along with the consideration of different complex solar irradiance patterns for validating its performance under different complex practical scenarios. This highlights the importance of the strategical study and in-depth analysis of the solar irradiance pattern for anticipating the envelop of power generation limit by the Solar PV system. The accurate forecasting of solar irradiance is necessary for enhancing the energy management and mitigating the performance degradation due to potential weather-related disruptions. In order to develop the accurate and effective solar irradiance forecasting model, the thesis delves into the exploration of different Neural Network (NN) models including ANN, Cascade Neural Network (Cas-NN) and Recurrent Neural Network (RNN). Additionally, the concepts of techniques of data pre-processing have also been integrated with these models for handling their computational burdens. The data pre-processing techniques have often been used for normalizing, cleaning, or modifying the dimensionality of the original input datasets. The advanced data pre-processing techniques like feature selection, feature clustering, etc. have also been employed in the thesis for restructuring the input dataset which makes the dataset more informative. The further in-depth analysis of these models has been done by integrating metaheuristic optimization algorithms for optimizing the internal parameters. Therefore, by utilizing the concept of optimizing the internal parameters of the forecasting model, a novel solar irradiance forecasting model, kPCA-HHO-RNN, has been developed by integrating appropriate data pre-processing technique (kPCA) with HHO optimized NN models. To test the scalability and practicability of the proposed model, its performance has been validated on the diverse range of datasets of different geographical locations. In addition to the solar irradiance forecasting, the thesis also dealt with the domain of electric load forecasting which plays a critical role in ensuring the effective performance of the power system. With the help of the insights gained from the solar irradiance forecasting methodologies, the data driven approach has been used for developing robust electric load forecasting model. The optimization of internal parameters of the forecasting model has been observed as the useful tool for enhancing its forecasting accuracy. However, for the complex search spaces, the traditional metaheuristic optimization algorithms suffer with the limitation of unbalanced exploration and exploitation phases. This unbalance may lead to premature curve vi convergence due to the entrapment at the local optimal point of solution. Therefore, a novel improved version of HHO algorithm (iHHO) has been developed using the hunting and encircling characteristics of the Grey Wolf Optimization (GWO) algorithm. A sophisticated hierarchical improvement has been introduced in the traditional HHO algorithm for enhancing its effectiveness for the complex search spaces. Several unimodal and multimodal objective functions have been used for validating the performance of the proposed iHHO algorithm. Additionally, with the help of proposed iHHO algorithm, a novel electrical load forecasting model has been developed. The in-depth comparative analysis of the proposed iHHO-based electric load forecasting model with other traditional optimization algorithms-based models, suggests that iHHO based model achieves superior performance with higher accuracy and lower error magnitudes. Furthermore, the extended evaluation suggests that the proposed electric load forecasting model outperforms the other recently developed state-of-art methodologies reported in the literature.