Development of forecasting models for photovoltaic energy systems: Data mining and deep learning approaches

PV energy systems are one of the leading renewable energy systems worldwide. Accurate forecasting of PV energy production is crucial for maintaining efficient energy grids, making informed decisions in the energy market and reducing maintenance costs. Solar radiation and temperature are two meteorological variables that directly affect the amount of PV energy produced in PV plants. In the first part of this thesis, a comprehensive model for predicting solar irradiance and temperature is built using various machine learning methods including IBK, LR, RF, RT, MLP and MLP Regression on power plant and meteorological station data from Harran University SPP and GAPYENEV Center. Then, for time series analysis, forecasting models were built using SES, ETS, HOLT, ARIMA and HWES. For both solar irradiance and temperature, daily forecasts as well as 1-step forward forecasts at 5-minute intervals were performed. It is crucial to subject both variables to the same methodology to create accurate models for PV forecasting. The generated energy yield obtained from the solar irradiance and temperature forecast values is compared with the actual energy yield of the PV power plant. PR is a critical parameter for evaluating PV plant performance, so PR calculation was performed. A detailed analysis of all machine learning and time series prediction models is presented by comparing them according to statistical error criteria such as RMSE, MAPE, MABE, MAE, MSE and DAC. In the second part, solar irradiance prediction models were created with deep learning. As the amount of collected renewable energy data increases, deep learning is becoming increasingly important in obtaining the needed big data analysis solutions. Therefore, this thesis proposes a hybrid model for solar irradiance prediction based on deep learning including CNNand LSTMnetwork and applies prediction models on existing PV systems. Furthermore, the predictive ability of the proposed models is evaluated in comparison with standard machine learning methods such as SVRand FFNN. Statistical error metrics such as RMSE, MAPE, MAE and R2 are used to measure the predictive ability of the models.