Enhancing Irradiance Prediction Using Short-Term Data: A Machine Learning Approach to Resilient Energy Management

Accurate irradiance prediction is necessary for the effective integration of photovoltaic (PV) systems into smart grids. However, rapidly changing weather patterns pose a challenge, rendering long-term historical data less reflective of current weather trends. This study, therefore, investigates the use of short-term data (Oct’22 - Apr’23) for hourly irradiance prediction at Kengeri, Bengaluru. Multiple Linear Regression (MLR) is employed as the base case. Subsequently, Random Forest (RF), Estreme Gradient Boosting (XGBoost), and a combination of Isolation Forest (IF) with RF and XGBoost are applied to enhance prediction accuracy. The results indicate cloud cover and humidity as key weather parameters influencing irradiance. The prediction improved upon the addition of hour polynomials and irradiance lag values to the weather parameters. For the MIR model, the R2, R.MSE, and M.AE value were 0.86, 111.19 W/sqm, and 83.62 W/sqm, respectively. In contrast, the most suitable model (IF+RF) model achieved an R 2 of 0.90, an RMSE of 94.51 W/sqm, and an MAE of 61.24 W/sqm, marking significant improvement against the baseline. It achieved a 24% reduction in overprediction and 31% in underprediction, with a daily average irradiance prediction error of −337.15 W/sqm. A consistent trend of underpredicting is observed across all models, struggling to accurately account for unpredictable intra-day cloud cover fluctuations due to limited variations in the weather conditions. Future work would extend the observation period and explore more complex models. Improvements in irradiance predictions are vital to optimize the operation of PV systems, supporting smart grid integration and a more resilient, sustainable energy infrastructure.