Cross-Plant Evaluation of Machine Learning and Deep Learning Models for Short-Term Photovoltaic Power Forecasting

摘要

The prediction of photovoltaic (PV) power output for short periods serves as a vital component which maintains grid stability while assisting with operational planning and renewable energy system power distribution. The research evaluates traditional machine learning techniques and deep learning techniques to determine their effectiveness in forecasting photovoltaic energy production during 15-minute intervals. The study develops four models which include Linear Regression (LR), Decision Tree (DT), Random Forest (RF), and Long Short-Term Memory (LSTM) to test their performance through the same testing process. The multivariate forecasting structure requires six input features which include historical yield data, ambient temperature data, module temperature data, irradiation data, temperature difference data, and squared irradiation data. The research examines how different time periods affect results by using four sliding window sizes, which include 5, 12, 24, and 48. The models were developed using data from Plant 1 and their performance was evaluated by testing them on Plant 2 data which had not been used before. The test results use the coefficient of determination ( ), mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), and training time to evaluate performance. The research findings demonstrate that LR has the highest accuracy performance across every window dimension because it produces results of 0.9592, 0.9606, 0.9563, and 0.9493 with minimal processing requirements. The LSTM model demonstrates its ability to compete in performance during short time windows, although its accuracy drops while processing power requirements rise with increased window dimensions. The RF algorithm delivers strong results, but the DT algorithm produces the weakest performance results. The research findings show that traditional machine learning systems provide better prediction accuracy and lower computational costs than the advanced LSTM system for the dataset and forecasting situation being studied.