Power Prediction of Photovoltaic System using Neural Network Models

This work considers a photovoltaic (PV) system installed on the rooftop of Agder Energi’s headquarters located at Kjøita, Kristiansand. The system includes three different types of solar PV modules; Suntech (multi-Si), Sharp (a-Si/μ-Si) and REC (multi-Si), that have a total installed DC capacity of 45 kWp. The system is grid-connected and instrumented for research and monitoring purposes. Artificial Neural Network (ANN) models were trained to obtain the lowest mean square error (MSE), by testing different configurations using a model-based trial and error approach. The model configurations that gave the lowest (MSE) were used to predict the power production from each of the PV modules using forecasted weather parameters obtained from MEPS (MetCoOp Ensemble Prediction System), with a one-day ahead and two-days ahead forecast horizon. The input selection of the models was based on both model-free and model-based approach, where the final input selection resulted in global horizontal irradiance, wind speed, air temperature and air mass, with the power (AC) production as output (target). The results indicated that the model configurations of 20 hidden neurons in first hidden layer, and 2 hidden neurons in second hidden layer gave the lowest MSE for all PV modules. Results from the test sets showed that the best model for Suntech gave MSE = 0.0454, Sharp gave MSE = 0.0325 and REC gave MSE = 0.0492. R2-values between 0.95 and 0.96 were obtained for all three models, indicating good fitting of the predicted values and the targets. Testing the Suntech and REC models with a hold-out set provided slightly less precise predictions compared to the results from the test set, while a higher precision was found for Sharp modules. Testing the model configurations with forecasted weather parameters indicated that the forecast accuracy of the weather will influence the power prediction, and the performance parameters will be accordingly. The one-day ahead forecasts provided MSE equal to 0.2647, 0.2378 and 0.2647, and for the two-days ahead forecast horizon an MSE equal to 0.2996, 0.2252 and 0.2719 for Suntech, Sharp and REC, respectively. An error much higher compared to the test set and hold-out set for the models, which inevitably was expected due to the weather forecast uncertainties. Based on the findings in this work, it can be concluded that a further optimization of the models will be necessary before obtaining even more precise predictions. However, the models did show good fitting for several days and a potential for using ANN models for power prediction of PV modules.