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

The study addresses the need for accurate forecasting of solar PV plant output to improve grid management and ensure economic sustainability in the face of renewable energy intermittency. With growing integration of distributed energy resources, AI techniques can enhance monitoring, control, and prediction of PV performance by leveraging sensor and IoT data. Weather significantly influences PV generation, making predictive modeling crucial for demand–supply balancing. The research identifies a gap: prior works often focus on solar irradiation prediction or use a limited set of AI methods and inputs, with few studies modeling both energy yield and performance ratio (PR) using multi-year real plant data. Objectives are to: (1) model energy yield and PR using RSM, ANN, and ANFIS with monthly tilted irradiation (MTI), wind speed (WS), and air temperature (AT) as inputs; (2) compare AI predictions with actual plant performance; and (3) validate models using Taylor’s diagram and statistical indices (R, R^2, NSCE, MAPE, RMSE, KGE, Theil’s U2).

Most prior research emphasizes solar radiation prediction rather than direct PV electricity generation, which depends on both hardware and meteorological factors. Studies have used ANFIS, ANN, SVM, numerical regression, and RSM for PV-related predictions. Examples include: SVM-based weather categorization for a 20 MW PV plant with ~8.46% prediction error using day type as input; RSM to model PV system thermodynamic and exergetic outputs; comparative works where RSM slightly outperformed ANN for radiation prediction. ANN-based works predicted short-term generation with varying accuracy (errors from ~4–31%). Deep learning approaches (LSTM, autoencoders) have shown advantages over MLP and physical models in some settings. Hybrid fuzzy–NN approaches (ANFIS) outperformed conventional statistics in case studies. Other efforts highlight the importance of including more weather parameters and integration with smart grid forecasting (including wind speed). Overall, comprehensive comparisons of multiple AI tools for both PV power and PR using multi-year real plant data remain limited, motivating this study.

Study site and data: A 2 MWp grid-connected PV plant at Kuzhalmannam, Kerala, India (ANERT) with SRRA station provided three years (2018–2020) of monthly operational and meteorological data via an integrated SCADA system (IEC 61724). Sensors included a pyranometer, 3-cup anemometer, and temperature sensors. Inputs and outputs: Inputs were monthly tilted irradiation (MTI, kWh/m^2), wind speed (WS, m/s), and ambient air temperature (AT, °C). Outputs were power generation (PG, kWh) and performance ratio (PR, %). Data preprocessing: Pearson correlation analysis indicated strong correlation between MTI and generation (R=0.9116), moderate between AT and generation (R=0.5264), and lower correlations for PR (AT to PR R=0.1747). Modeling approaches: (1) RSM: Built empirical polynomial models (including interaction and higher-order terms). A cubic model was fit for PG and a transformed (sqrt[PR]) cubic model for PR. ANOVA assessed term significance; model forms included linear, interaction, quadratic, and cubic terms. (2) ANN: Multilayer feed-forward network with 3-10-2 architecture (3 inputs, one hidden layer with 10 neurons, 2 outputs), trained with Levenberg–Marquardt (trainlm). Data split: 70% training, 15% validation, 15% testing. Performance measured by MSE and statistical indices. (3) ANFIS: First-order Sugeno-type MISO models for PG and PR using grid partitioning to define membership functions and fuzzy rules; hybrid learning algorithm in MATLAB (2016). Data split: 70% training, 30% validation. Model evaluation: Statistical indices included Pearson’s R, R^2, NSCE, RMSE, MAPE, KGE, and Theil’s U2, plus Taylor diagrams to visualize correlation, standard deviation, and centered RMSE relative to observations. Response surfaces and contour plots analyzed input–output interactions.

- Data correlations: MTI strongly correlated with generation (R=0.9116); AT had moderate correlation with generation (R=0.5264) and weaker with PR (R=0.1747). WS showed weak negative correlations with both outputs. - RSM results: - Power generation (PG): R=0.9886, R^2=0.9773, NSCE=0.9774, MAPE=2.24%, RMSE=6133.93, KGE=0.9847, Theil’s U2=0.0775. Response surfaces indicated non-linear interactions: at lower MTI, PG initially rises with WS then declines; at higher MTI, PG decreases then increases with WS; AT showed overall positive influence. - Performance ratio (PR): R=0.9346, R^2=0.8735, NSCE=0.8738, MAPE=2.05%, RMSE=1.85, KGE=0.9157, Theil’s U2=0.3343. PR increased with higher MTI; PR peaked at mid-range WS; high AT combined with high WS improved PR. - ANN results: - PG: R=0.9679, R^2=0.9369, NSCE=0.9128, MAPE=3.77%, RMSE=12070, KGE=0.9096, Theil’s U2=0.325. - PR: R=0.9663, R^2=0.9337, NSCE=0.9317, MAPE=1.5%, RMSE=1.37, KGE=0.9638, Theil’s U2=0.245. - ANFIS results: - PG: R=0.9950, R^2=0.9901, NSCE=0.9828, MAPE=2.09%, RMSE=5492.81, KGE=0.956, Theil’s U2=0.1506. - PR: R=0.9915, R^2=0.9830, NSCE=0.9837, MAPE=0.8%, RMSE=0.6898, KGE=0.9917, Theil’s U2=0.1259. - Comparative performance (Taylor diagrams and indices): ANFIS consistently outperformed ANN and RSM for both PG and PR, showing higher correlation with observations, lower errors, and lower uncertainty. ANN was second-best overall; RSM offered interpretable polynomial models with reasonable accuracy. - Practical implication: Using only three readily measurable meteorological inputs (MTI, WS, AT), high-accuracy forecasting of PV energy yield and PR is achievable, with ANFIS providing the most precise results in this case study of a 2 MWp plant in Kerala, India.

The findings demonstrate that AI-based models can effectively capture the nonlinear relationships between weather variables and PV plant performance. By comparing RSM, ANN, and ANFIS using identical inputs and multi-year real plant data, the study shows that ANFIS offers the strongest agreement with observed energy yield and PR, indicating superior ability to model complex interactions and uncertainties inherent in monthly aggregated PV performance. The strong performance of ANFIS addresses the research objective of identifying an accurate predictive tool for PV output and PR, supporting improved grid planning, energy dispatch, and operational decision-making. RSM provides transparent mathematical relationships valuable for interpretation, while ANN delivers competitive accuracy but with slightly higher errors and uncertainty than ANFIS. The response surface analyses further clarify how MTI, WS, and AT jointly influence outputs, informing performance optimization and expectation management across varying climatic conditions.

AI-driven prediction models for PV performance were developed and validated using three years of operational data from a 2 MWp plant. Among the compared approaches (RSM, ANN, ANFIS), ANFIS delivered the highest predictive accuracy for both power generation and performance ratio, evidenced by the highest R^2 and NSCE values and the lowest errors and uncertainty. RSM enabled interpretable polynomial relationships and ANN provided competitive results, but both were outperformed by ANFIS. Accurate forecasting of PV output using a minimal set of meteorological inputs can assist utilities, policymakers, and developers with grid integration, scheduling, and investment planning. The study underscores the practicality of data-driven models for performance mapping of utility-scale PV plants.