Weather Impact on Solar Farm Performance: A Comparative Analysis of Machine Learning Techniques

Renewable energy is vital for mitigating climate change. The intermittent nature of renewable sources necessitates improved forecasting and modeling for effective grid management. Addressing supply chain vulnerabilities and incorporating storage systems enhances the reliability of large-scale solar power. Artificial intelligence (AI) plays a crucial role in integrating renewables into the grid and improving their competitiveness. AI, coupled with sensors and IoT devices, provides valuable insights for grid operators, enabling dynamic control and optimization of loads, and resource selection based on availability. This leads to more intelligent microgrids and virtual power plants. Weather significantly impacts renewable energy generation, making accurate prediction of solar and wind power output essential for global electricity demand and supply management. AI techniques offer the potential for precise PV plant performance prediction, enhancing efficiency and accessibility, and addressing the variability inherent in renewable energy generation. AI's capacity to learn complex patterns from data eliminates the need for complex mathematical routines and rules, making it a powerful tool in this context. The increasing reliance on distributed energy resources further strengthens the case for AI integration in grid management and optimization of renewable energy resources.

Existing research primarily focuses on predicting solar radiation, with limited studies on predicting PV electricity generation. While solar irradiation is a key factor, other parameters like hardware specifications (cell size, type, incidence angle, layout) and weather conditions (temperature, wind speed, relative humidity) also influence power output. Several researchers have used AI tools such as ANFIS, ANN, numerical regression, support vector machines (SVM), and RSM to predict PV power generation based on weather data. Studies have shown varying levels of prediction accuracy using these techniques. For example, Shi et al. [14] used SVM with a prediction error of 8.46%, while Kazemian et al. [15] employed RSM for photovoltaic system modeling and optimization. Other studies have employed ANN [17], LSTM [18, 19], and hybrid approaches combining fuzzy logic and neural networks [20] with varying degrees of success. However, many studies focus on solar radiation prediction or use only one or two AI methods. A comprehensive comparison of multiple AI techniques for PV plant performance modeling, considering various performance metrics and a substantial dataset, is lacking. This study aims to address this gap by using three years of real-time PV plant data to model and compare the performance of RSM, ANN, and ANFIS techniques in predicting both power generation and performance ratio (PR).

This research utilizes data from the 2 MW Kuzhalmannam Solar PV Project in Kerala, India, which has been operational for over four years. The project features a Solar Resource Assessment (SRRA) facility that gathers meteorological data, and a Supervisory Control and Data Acquisition (SCADA) system collects real-time data on solar irradiance, energy production, wind speed, ambient temperature, and module temperature. Three years of data (2018-2020) were used for model development and validation. Three input parameters were selected: monthly tilted irradiation (MTI), wind speed (WS), and air temperature (AT). Two output parameters were considered: energy yield (cumulative energy generation per month) and performance ratio (PR). Data preprocessing involved analyzing correlations between input and output variables using correlation heatmaps and matrices. Three AI techniques were employed: Response Surface Methodology (RSM), Artificial Neural Network (ANN), and Adaptive Neuro-Fuzzy Inference System (ANFIS). RSM involved developing mathematical models (first and second-order) to represent the relationship between inputs and outputs. ANN used a multilayer feed-forward neural network with one input layer, one hidden layer (number of neurons determined through trial and error), and one output layer. The Levenberg-Marquardt algorithm was used for training. ANFIS combined fuzzy logic and neural networks, employing a Sugeno-based fuzzy inference system. Model performance was evaluated using statistical indices: Pearson's R, R², RMSE, MAPE, NSCE, KGE, and Theil's U². Taylor diagrams were used to visualize model performance and compare the different AI techniques. The dataset was divided into training, validation, and testing sets for model development and evaluation.

RSM, ANN, and ANFIS models were successfully developed for predicting both power generation and performance ratio. RSM generated mathematical equations representing the relationship between inputs and outputs. For power generation, RSM achieved an R² of 0.9773, RMSE of 6133.93, and MAPE of 2.24%. For the performance ratio, RSM achieved an R² of 0.8735, RMSE of 1.85, and MAPE of 2.05%. ANN models showed high correlation (R > 0.96) for both power generation and performance ratio prediction. The ANN model for power generation had an R² of 0.9369, RMSE of 12070, and MAPE of 3.77%. For performance ratio, the ANN model had an R² of 0.9337, RMSE of 1.37, and MAPE of 1.5%. ANFIS models exhibited the best performance in this study. For power generation, ANFIS achieved an R² of 0.9901, RMSE of 5492.81, and MAPE of 2.09%. For performance ratio, ANFIS reached an R² of 0.9830, RMSE of 0.6, and MAPE of 0.8%. Taylor diagrams visually confirmed ANFIS's superior performance in predicting both power generation and performance ratio compared to ANN and RSM. Theil's U² values indicated low uncertainty in the ANFIS model for both parameters. Across all models, Monthly Tilted Irradiation (MTI) was consistently identified as the most influential input variable for both power generation and performance ratio.

The findings demonstrate the effectiveness of AI in accurately predicting the performance of solar PV plants. ANFIS emerged as the superior model, achieving the highest accuracy in predicting both power generation and performance ratio, as indicated by multiple statistical metrics. The high R² values and low RMSE and MAPE values for ANFIS suggest its ability to capture the complex non-linear relationships between weather parameters and PV plant performance. The superior performance of ANFIS likely stems from its ability to effectively integrate the strengths of both fuzzy logic and neural networks, offering a robust and adaptable framework for modelling the intricate dynamics of PV systems. This study highlights the potential of using data-driven AI models for improving solar energy forecasting and optimizing grid management. The developed models provide valuable tools for policymakers and solar energy professionals to enhance the efficiency, sustainability, and profitability of solar energy systems.

This research successfully demonstrated the use of AI techniques (RSM, ANN, and ANFIS) to predict solar PV plant performance. ANFIS consistently outperformed ANN and RSM across multiple metrics, indicating its suitability for accurate power generation and performance ratio forecasting. This study's findings are valuable for improving solar energy forecasting and grid management. Future work could explore the integration of additional weather parameters, longer time series data, and more sophisticated AI algorithms to enhance predictive accuracy and explore the applicability of these models to different geographic locations and PV technologies.

The study's findings are based on data from a single 2 MWp PV plant in Kerala, India. The generalizability of the models to other geographic locations with different climate conditions and PV technologies needs further investigation. The study also focused on a limited set of input parameters. Including additional parameters such as humidity, cloud cover, and solar azimuth angle might improve model accuracy. The three-year data period may not fully capture long-term variations in weather patterns and PV plant performance. Finally, the computational cost of training the ANFIS model could be a concern for very large datasets.