Predictive Modeling and Management of Solar Power Generation in a Factory: A Case Study in the Indian Region

This paper investigates the prediction of solar power generation using three machine learning prediction models: ARIMA, SARIMAX, and PROPHET. The study collected a dataset of 34 days from a solar power plant in a region of India, including generation data at each inverter level and sensor readings at the plant level. Sensor and inverter level issues were identified after checking the data integrity and EDA. Erroneous data was eliminated and then introduced into the model to improve the stability of the predictions. The predictions are more accurate since the SARIMAX model is based on the ARIMA model with seasonal factors associated with solar power generation. The data were introduced into SARIMAX and PROPHET models for computational analysis and comparison. The results show that the R2 Score of ARIMA is 0.894025, the MAE Score is 15882.183111, and the RMSE Score is 25400.796426, while the scores of the above three evaluation indexes of the PROPHET model are 0.897808, 20236.216888, and 27049.765423 respectively, which indicates that the accuracy of the SARIMAX model is slightly higher than that of the PROPHET model. However, the PROPHET model is easier and faster to use. Compared with the traditional prediction methods, both models can achieve highly accurate PV power prediction at a lower cost and have some ability to deal with complex nonlinear problems. The above study shows that the SARIMAX and PROPHET models are highly accurate and easy to use for solar power prediction and can provide an adequate prediction method for the solar power industry.