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

Renewable energy is the optimal solution for mitigating the threats of climate change. The renewable energy sector has made significant progress due to technological advancements. However, the intermittency of renewable energy sources necessitates improved forecasting and modeling of power resources for effective grid management. Addressing the vulnerabilities in renewable energy supply chains is crucial to manage variability. Incorporating storage systems benefits large-scale solar power developments, and intelligent systems, particularly those integrating artificial intelligence (AI), can enhance the integration of renewables into the grid and improve market competitiveness. AI integration, enabled by sensors and IoT devices, provides valuable insights for grid operators. Hybridization and storage are also gaining popularity with solar PV plants, enhancing grid stability and reliability. The growing adoption of distributed energy resources requires AI techniques to optimize loads and manage renewable energy resource selection based on availability. AI integration makes microgrids and virtual power plants more dynamic and intelligent. AI can significantly improve solar power plant performance, with weather conditions being a major influencing factor on solar and wind plant energy generation. Accurate prediction of wind and solar PV plant output is essential for electricity system demand and supply management globally. AI techniques offer high potential for weather and renewable energy performance prediction, learning critical patterns without complex mathematical routines or rules. Intelligent sensors and IoT systems are interconnected to collect substantial amounts of data for analysis and prediction.

Existing research primarily focuses on predicting solar radiation, with limited work on forecasting PV electricity generation. While solar irradiation is a key factor, other elements like hardware (cell size, solar cell type, incidence angle, layout) and weather conditions significantly impact power output. For instance, solar cell temperature influences electricity generation, affected by solar irradiation, ambient temperature, wind speed, and relative humidity. Researchers have used various AI tools, including ANFIS [9], ANN [10], numerical regression [11], support vector machines [3, 12], and RSM [13], often based on weather categorization, to predict PV power generation. Studies have explored different approaches, such as support vector machines for weather categorization to estimate power production [14] (with prediction error of 8.46%), RSM for photovoltaic system prediction [15], and comparing ANFIS, ANN, and RSM for solar radiation prediction [16]. Other research has employed ANN for predicting daily power generation [17], explored deep learning techniques like LSTM [18, 19] and autoencoders for improved power prediction, used a hybrid approach of fuzzy decision and neural networks [20], and applied LSTM for predicting energy production from large solar plants, considering uncertainty and weather forecasts [21]. The literature also includes studies on neural network-based fault diagnosis in PV installations [22], ANN models for daily solar radiation estimation for power capacity estimation [23], and the impact of incorporating more weather parameters for enhanced prediction accuracy [24]. Deep learning methods for short-term wind speed prediction have also been investigated [25, 26]. However, a comprehensive comparison of various AI-based prediction tools for PV plant performance modeling, evaluation, and metrics using extensive, real-time data over multiple years has been lacking.

This research employs three AI techniques (RSM, ANN, ANFIS) to predict the performance of a 2 MWp solar PV plant in Kuzhalmannam, Kerala, India, using three years (2018-2020) of real-time data. The study uses monthly tilted irradiation (MTI), wind speed (WS), and air temperature (AT) as input parameters, and energy yield and performance ratio (PR) as output parameters. Data was collected using a Supervisory Control and Data Acquisition (SCADA) system integrated with pyranometers, anemometers, and temperature sensors. Data pre-processing involved correlation analysis (heatmap and correlation matrix) to assess the relationship between input and output variables. The predictive models were evaluated using statistical indices: Pearson's R, R², NSCE, RMSE, MAPE, and KGE. Theil's U² and Taylor's diagram were used to assess prediction uncertainty. **RSM Methodology:** RSM, a combination of statistical and mathematical approaches, was used to design trials for accurate response prediction, model fitting, and parameter optimization. A quadratic or cubic model (Equation 1 and 2) was used to approximate the response (Equation 3). For power generation and PR, separate RSM models were developed (Equations 13 and 14), and statistical indices were calculated to evaluate model performance. **ANN Methodology:** A multilayer feed-forward ANN with one input layer (three neurons), one hidden layer (ten neurons determined through trial and error), and one output layer (two neurons for PG and PR) was used. The Levenberg-Marquardt (trainlm) training function was employed, and the data was divided into training (70%), validation (15%), and testing (15%) sets. Model performance was evaluated using MSE, R, R², MAPE, NSCE, KGE, and Theil's U². **ANFIS Methodology:** An ANFIS model was developed using a first-order Sugeno model with three input variables (MTI, WS, AT) and one output variable (either PG or PR). MATLAB 2016 was used for model development and training using a hybrid learning technique. A grid partitioning approach was used for FIS structure creation, and the centroid defuzzification method was employed. Model performance was evaluated using the same statistical indices as the ANN model.

All three AI models (RSM, ANN, ANFIS) effectively predicted power generation and performance ratio, showing high correlations between predicted and actual values. However, ANFIS demonstrated superior performance. **Power Generation:** * RSM: Achieved R = 0.9886, R² = 0.9773, NSCE = 0.9774, MAPE = 2.24%, RMSE = 6133.93, KGE = 0.9847, Theil's U² = 0.0775. * ANN: Achieved R = 0.9679, R² = 0.9369, NSCE = 0.9128, MAPE = 3.77%, RMSE = 12070, KGE = 0.9096, Theil's U² = 0.325. * ANFIS: Achieved R = 0.9950, R² = 0.9901, NSCE = 0.9828, MAPE = 2.09%, RMSE = 5492.81, KGE = 0.956, Theil's U² = 0.1506. **Performance Ratio:** * RSM: Achieved R = 0.9346, R² = 0.8735, NSCE = 0.8738, MAPE = 2.05%, RMSE = 1.85, KGE = 0.9157, Theil's U² = 0.3343. * ANN: Achieved R = 0.9663, R² = 0.9337, NSCE = 0.9317, MAPE = 1.5%, RMSE = 1.37, KGE = 0.9638, Theil's U² = 0.245. * ANFIS: Achieved R = 0.9915, R² = 0.9830, NSCE = 0.9837, MAPE = 0.8%, RMSE = 0.6898, KGE = 0.9917, Theil's U² = 0.1259. Taylor diagrams visually confirmed ANFIS's superior performance for both power generation and performance ratio prediction, showing its points closer to the baseline than ANN and RSM.

The study successfully demonstrates the effectiveness of AI techniques, particularly ANFIS, in predicting solar PV plant performance. ANFIS consistently outperformed ANN and RSM in terms of accuracy and low prediction uncertainty, as evidenced by the statistical indices and Taylor diagrams. The high accuracy of ANFIS suggests its potential for practical application in forecasting solar energy generation, aiding policymakers and solar energy researchers in making informed decisions regarding grid integration and resource management. The use of three years of real-time data enhances the robustness and reliability of the findings compared to previous studies relying on shorter datasets or simulated data. However, further investigation is needed to assess the generalizability of these findings to different geographical locations and climatic conditions.

This research successfully developed and compared three AI models (RSM, ANN, and ANFIS) for predicting solar PV plant power generation and performance ratio. ANFIS consistently outperformed the other models, demonstrating superior accuracy and lower uncertainty. This highlights the potential of ANFIS for precise solar energy forecasting, valuable for grid management and resource planning. Future research should focus on model validation in diverse climatic zones and explore more advanced machine learning algorithms to further enhance prediction accuracy and incorporate factors like performance degradation over time.

The study's findings are based on data from a single 2 MWp solar PV plant in a specific location (Kuzhalmannam, Kerala, India). This may limit the generalizability of the results to other PV plants with different characteristics or installed in various geographical locations and climatic conditions. The models' accuracy might also be influenced by the quality and completeness of the input data, highlighting the importance of reliable data acquisition and preprocessing steps. Future research should consider expanding the dataset to encompass a wider range of PV plant characteristics, geographical locations, and climatic conditions to enhance the generalizability and applicability of the proposed models.