| Abstract [eng] |
This master's thesis examines the development of photovoltaic power generation forecasting models and their integration into an industrial power plant control system. The relevance of the work is associated with the increasing share of renewable energy in the electrical power system and the need to accurately estimate, in real time, the maximum available power of a photovoltaic power plant. This value is essential for power curtailment, reserve maintenance, compliance with grid operator requirements, and reliable plant control. The thesis presents a review of photovoltaic power forecasting models, covering physical, analytical, statistical, machine learning, and hybrid methods. In the experimental part, several forecasting models were developed and compared: MLP, RF, ARIMA, RNN, LSTM, CNN, and hybrid structures. The models were evaluated using MAE, RMSE, and MSE metrics, as well as by graphical comparison of predicted power curves with actual active power and the currently used Laplace transform-based model in the power plant. The study found that although some models achieved high accuracy under certain conditions, their implementation in a PLC controller was more complex due to a large number of parameters, state storage requirements, or sequence processing needs. The MLP model was selected as the best compromise between accuracy, number of parameters, computational simplicity, and feasibility of implementation in the CODESYS ST environment. An important outcome of the work is the developed IoT infrastructure for exporting, transmitting, and applying model parameters in a real control system. Model parameters are transmitted to the PLC via an MQTT interface, where they are decoded and used for real-time active power prediction. In addition, a containerized automated model training system was developed, operating in a Linux VM environment. This system automatically collects data from a time-series database, prepares training datasets, filters curtailed generation states, trains models, and selects the best-performing one. The final MLP system achieved RMSE = 540.331 kW and MAE = 157.201 kW on the test dataset. Real power plant operation tests demonstrated that the MLP model running on the PLC controller can achieve lower errors than the currently used Laplace transform-based model. The MLP model reduced MAE by 41.50% and RMSE by 35.37%. These results confirm that the developed model and its integration infrastructure can be successfully applied in a real photovoltaic power plant control system. |
