Forecasting Behind the Meter Solar Energy Production using Machine Learning Algorithms

Solar power is one of the most prevalent renewable resources offered by nature, and the Sun's energy can be harnessed in various forms. Among them, the production of electrical power through solar photovoltaic (PV) systems has become one of the brightest and greenest alternatives. Nevertheless, environmental conditions such as irradiance and temperature have a significant impact on the behavior of solar PV systems, which are naturally variable and random in nature. To do the proper solar forecasting – first forecasting of temperature and irradiance are required. Therefore, accurate prediction of PV power generation is necessary for efficient system control, and load synchronization. The purpose of this study is to develop efficient and accurate solar PV power forecasting models using machine learning and deep learning techniques. And then I collected the data from Hongkong university of Science & Technology for the year 2023. Based upon this historical data it depicts the solar irradiance, and temperature is low at the beginning and at the end of the year. Through the experiment of LSTM and CNN algorithm I found that the LSTM model has better accuracy. For temperature prediction LSTM model used an epoch of 30 and CNN model used an epoch of 17. In contrast to the prediction of irradiance LSTM used an epoch of 29 and CNN used an epoch of 55. LSTM outperforms CNN for both temperature and irradiance forecasting. In temperature forecasting, LSTM recorded lower measurement errors — RMSE = 1.7037°C, MAE = 1.2431°C, and MSE = 2.9026°C — compared to the CNN model's RMSE = 4.5053°C, MAE = 3.4203°C, and MSE = 20.2980°C. Furthermore, LSTM's R² value of 0.8912 is significantly greater than CNN's R² = 0.2389 and hence more appropriate. Also in irradiance forecasting, the LSTM model again performs better with RMSE = 63.2078 W/m², MAE = 36.4779 W/m², and MSE = 3995.2312 W/m² than CNN's RMSE = 67.3843 W/m², MAE = 37.4451 W/m², and MSE = 4540.6452 W/m². The R² value of the LSTM model of 0.9412 is greater than 0.9332 for CNN, again confirming that the LSTM model has a better fit and more accurate predictions overall. Based on this model, one can see that it is possible to obtain very accurate solar forecasting, but temperatures could be very unstable and the weather unreliable, thereby sometimes making accurate energy demand forecasts could be difficult. By using this model, the utility companies can decide for those periods that experience high demand, control the generation of energy, and provide supply at a specific time.