Forecasting of produced output electricity in photovoltaic power plants

Abstract

The global energy market is pivotal to modern economic growth, with renewable energy, particularly solar power, emerging as a key strategy for mitigating climate change and fostering sustainable development. Solar energy, a clean and abundant resource, directly supports UN Sustainable Development Goals (SDGs) such as Goal 7 (Affordable and Clean Energy) and Goal 13 (Climate Action) while indirectly contributing to Goals 8 (Decent Work and Economic Growth), 9 (Industry, Innovation, and Infrastructure), and 11 (Sustainable Cities and Communities). However, the variability of solar energy generation—shaped by factors like solar radiation, temperature, and technological efficiency—poses challenges for grid stability, resource planning, and the broader integration of renewables. Accurate forecasting is critical to addressing these issues, enabling better grid management, informed policymaking, and strategic investments in renewable energy systems. This thesis tackles the challenges associated with solar energy forecasting by utilizing cutting-edge developments in artificial intelligence (AI) and machine learning (ML) to improve prediction accuracy and reliability. It focuses on under-explored areas, including the integration of endogenous and exogenous variables, chaotic modeling, and hybrid model optimization. Through systematic investigation and evaluations, the research proposes innovative techniques validated with real-world data, providing practical insights for managing solar power generation and advancing low-carbon energy systems. By tackling critical gaps in forecasting methodologies, this work supports global sustainability goals, promotes energy security, and accelerates the transition to cleaner and more efficient power networks. This thesis builds upon the extensive literature review to identify and address critical gaps in solar energy forecasting. The research focuses on four key dimensions: (1) integrating both endogenous variables and exogenous variables for a holistic understanding of solar energy generation, (2) employing robust data preprocessing techniques like dimensionality optimization and feature extraction to enhance data quality and model reliability, (3) exploring advanced deep learning architectures and ensemble frameworks to improve predictive performance, and (4) optimizing metaheuristic algorithms for efficient and scalable AI-based forecasting solutions. These four methodological pillars aim to address challenges such as nonlinearity, temporal dependencies, and chaotic behavior in solar power datasets. The study employs five groups of forecasting models—linear, machine learning, deep learning, ensemble learning, and chaotic models—validated using real-world data from the Konya Eregli Solar Power Plant. Among the models, Feedforward Neural Networks (FFNN), Extreme Learning Machines (ELM), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Echo State Networks (ESN) emerged as top performers based on standard evaluation metrics. These seven models were tested under scenarios combining endogenous and exogenous inputs and underwent hyperparameter optimization using the Tree-structured Parzen Estimator (TPE) to maximize accuracy. Together, these advancements contribute to more reliable and precise solar energy forecasting, supporting the integration of renewable energy into power grids. Prior to training the models, the dataset undergoes a thorough cleaning process to ensure the quality and reliability of the data. Outliers are identified and removed using the Local Outlier Factor (LOF) method, followed by an exploratory data analysis to gain deeper insights into the time series characteristics. Once the data is cleaned and understood, feature engineering strategies are applied to extract and transform relevant features. These strategies are specifically tailored to meet the unique requirements of both individual base models and ensemble stacking meta-models. To ensure robust model evaluation and prevent data leakage, the processed dataset is split into training, validation, and test sets, forming a reliable foundation for both base and meta-model development. This study evaluates the performance of all implemented forecasting models using five well-established evaluation metrics: R-squared (R2), Root Mean Square Error (RMSE), Mean Squared Error (MSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). Together, these metrics offer a thorough analysis of the models' accuracy and reliability. Among the fine-tuned base models, the LSTM and GRU models demonstrate the highest levels of accuracy and reliability. These deep learning architectures effectively capture the complex temporal dependencies inherent in solar power data. Following LSTM and GRU, other models such as FFNN, XGBoost, RF, ESN, and ELM also perform well, with their efficacy varying based on specific scenarios. These results underscore the strengths of deep learning models while validating the applicability of traditional machine learning and ensemble methods in addressing particular forecasting challenges. To address chaotic behavior in the time series, the False Nearest Neighbors (FNN) method is incorporated as a preprocessing technique for reconstructing the phase space. This step determines an optimized embedding dimension intended to maintain the structural consistency of the original time series. The ESN algorithm is then employed to forecast solar energy production using both the original and embedded data. Results show that optimizing the embedding dimension via the FNN technique significantly improves the model's capability to capture the nonlinear dynamics underlying solar power generation, leading to more accurate forecasts. The enhanced performance of the ESN model on the embedded data underscores the vital role of data preprocessing—particularly embedding dimension selection—in chaotic modeling for solar energy systems. Overall, this approach demonstrates how combining proper data embedding with an optimized model architecture leads to superior predictive capabilities, further contributing to efficient and reliable solar energy forecasting. As the final stage of this thesis, stacking ensemble machine learning techniques are applied to enhance forecasting performance even further. The primary aim is to develop a novel metamodel that surpasses the prediction capabilities of the individual base models. Two categories of stacking ensemble models are investigated: regression-based stacking and neural network-based stacking. Each approach is critically evaluated and compared, leading to the identification of the most accurate model—an essential contribution of this research. In the first category, traditional regression algorithms—Multi-Linear Regression, Ridge Regression, and Lasso Regression—function as the meta-model to combine the predictions of the seven base models. This approach leverages the strengths of diverse base predictors through a linear combination, aiming to minimize overall forecast error. However, it ultimately does not surpass the established accuracy benchmark. In the second category, neural networks (NN) function as the meta-model, exploiting their capacity to capture complex nonlinear relationships within the data. Three distinct neural network-based stacking strategies are developed, including the proposed deep learning-based stacking model. This meta-model is deliberately constructed using only the two best-performing base models, selected based on their individual accuracy metrics. To further enhance predictive performance, advanced techniques and optimized hyperparameters are incorporated. As a result, this final stacking model delivers the highest forecast accuracy among all models investigated. The meta-models' designs undergo several refinements, including the integration of Model Output Statistics (MOS) as a post-processing technique. These adjustments aim to determine whether further tuning can enhance forecasting accuracy. The outcome of these efforts is the Deep Learning-Based Stacking Meta-Model, proposed by this thesis, which predicts solar energy output for the Konya Eregli Solar Power Plant with an impressive precision exceeding 98%. This model not only outperforms the most accurate base model (LSTM) but also achieves the primary objective of this research: setting a new standard for high-accuracy solar energy forecasting. A key strength of the proposed deep learning-based stacking meta-model lies in its simplified architecture, which stands out compared to the other meta-models developed in this thesis. By focusing exclusively on the two most accurate base models, the meta-model balances exceptional forecasting accuracy with reduced computational demands. This minimalistic design speeds up both the training and deployment processes, aligning with the research objective of optimizing metaheuristic model architectures to create more efficient and scalable AI-driven forecasting solutions. This level of predictive precision underscores the transformative potential of advanced ensemble techniques in solar energy forecasting. By achieving improved accuracy, these methods enhance energy management systems, enable smoother grid integration, and support strategic long-term planning in renewable energy sectors. As a result, the proposed deep learning-based stacking meta-model proves to be an essential tool for advancing sustainable energy infrastructures and ensuring a reliable transition to renewable energy systems.