Hourly Photovoltaic Power Forecasting Using CNN-LSTM Hybrid Model

Abstract

In addition to being environmentally friendly, Photovoltaic (PV) power generation has gained popularity globally because it is also abundant in nature. However, the stochastic nature of its radiative energy has made its use undesirable to some consumers at various levels. In this manuscript, a state-of-the-art hybrid deep learning model is introduced to forecast PV power on an hourly basis. The CNN-LSTM model is a hybrid of Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) models. This hybrid model was trained using historical data of Johannesburg city with two different data compositions. The first experiment involved training the model with 80% of data collected over one year from Johannesburg city. The second experiment involved training the model with 80% of ten-year historical data from the same city. Using Root Mean Square Error (RMSE) as the evaluation metric, the results obtained were compared with results from CNN and LSTM time series models. It was discovered that the CNN-LSTM model outperformed all the other models used in the same experiment. For instance, at 1,000 epochs and with ten years of data composition, the CNN-LSTM hybrid model gave an RMSE value of 7.21 Watts/m <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> while the LSTM and CNN time series models recorded 7.30 Watts/m <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> and 12.67 Watts/m <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> respectively. This outcome indicates that a hybrid of CNN and LSTM for short-term PV power forecasting performs better than when the models are used individually. It is suggested that deploying the CNN-LSTM model for short-term PV power forecasting in Johannesburg’s solar farms could be beneficial in stabilizing the PV power connection to the smart grid.