Machine Learning Based Photovoltaic Power and Energy Prediction in time-frequency domain

Abstract

Efficient energy management in smart-grid systems relies heavily on accurate photovoltaic (PV) power production forecasting. In this study, we explore the benefits of employing frequency domain methods for PV panel power production forecasting. moreover two methods are proposed in this article, Discrete Wavelet Transform with Long Short-Term Memory (DWT-LSTM) for short-term forecasting and LSTM combined with Short-Time Fourier Transform and an Artificial Neural Network (LSTM-STFT-ANN) for long-term forecasting. to assess the robustness of the proposed methods, extensive testing and comparison was conducted with the state of the art methods LSTM combined with a conventional neural network (CNN-LSTM) on an unknown dataset spanning two years, enabling prediction performance evaluation under varying weather conditions, including foggy and cloudy scenarios. Additionally, we evaluated the total energy produced across different time horizons. The performance assessment results for 1-hour horizon power forecasting revealed that the DWT-LSTM method achieved an average normalized root mean squared error (HH) of approximately 0.16, outperforming the classical CNN-LSTM method, which yielded an HH of more than 0.22. For longer-horizon forecasts, the LSTM-STFT-ANN method exhibited higher efficiency. For instance, for 2-day horizon, the HH indicator was recorded as 0.35 for LSTM-STFT-ANN, compared to around 0.54 for the CNN-LSTM method.