Solar radiation forecasting using machine learning techniques

This study focuses on the prediction of solar radiation for Diyarbakır province using machine learning techniques. Among renewable energy sources, solar energy has gained increasing importance and holds the potential to enhance energy efficiency with accurate prediction methods. In this context, a Neural Network (NN) model was developed using meteorological data collected at minute, hourly, daily, monthly, and yearly intervals from the meteorological station located on the central campus of Dicle University. During the data preprocessing phase, missing data were imputed, outliers were removed, and the dataset was normalized. The dataset was split into 90% training and 10% testing data, with 10% of the training data used for validation. The NN model consists of eight layers in total, including the input layer, and contains seven hidden layers. All hidden layers were trained using the ReLU activation function. The output layer consists of a single neuron and does not include any activation function. The Adam optimization algorithm was used with a learning rate of 0.001, and the model was trained over 1000 epochs. The results demonstrate that the model achieved successful prediction performance. The monthly average R² value was 0.849, while RMSE values ranged from 48.07 to 150.32 W/m², MSE values ranged from 2311.06 to 22596.75 W²/m⁴, and MAE values varied between 26.98 and 87.39 W/m². Accordingly, the prediction accuracy of the model varied by month, and the overall error metrics were calculated to fall within these specified ranges. The developed NN model was compared with other machine learning techniques such as Random Forest, Decision Tree, Support Vector Machines, and K-Nearest Neighbors, and it was observed to offer higher accuracy and lower error values. These results confirm that the NN model effectively captures complex and non-linear relationships, outperforming traditional empirical methods in generating more accurate predictions. The findings of this study highlight that accurate solar radiation predictions contribute significantly to critical processes such as selecting energy storage units, determining electricity production strategies, and optimizing the efficiency of solar panels.