A Compound of Deep-Learning and Feature Selection for Solar Power Forecasting Applications

The greenhouse gas effect escalates due to carbon emissions from traditional use of fossil fuel, particularly driven by the development and urbanization of the modern world. In this context, solar photovoltaic (Solar-PV) power generation offers a valuable opportunity to mitigate the impact of conventional fuels. However, accurate forecasting has consistently been a challenge in the context of Solar-PV systems. This paper presents an effort to address this issue by developing a new hybrid deep-learning method named as an improved WT-LSTM model. In first phase of the proposed methodology, the most relevant features that significantly impacted the power generation are selected using various feature selection methods, including Least absolute shrinkage and selection operator (LASSO) regression, coefficient of determination, forward selection, backward selection, and mutual information. Besides, in second phase, the Solar-PV power generation is predicted by applying the proposed improved WT-LSTM model. The proposed hybrid improved WT-LSTM model adopts a wavelet transform based decomposition of the historical input time series data of solar-PV power generation into various frequency components, from which statistical features are extracted. These statistical features are also combined with meteorological features and become the basis for Solar-PV generation forecasting using improved LSTM. Moreover, the proposed improved WT-LSTM model further incorporates batch normalization, dropout, and L2 regularization to forecast various time horizons. The algorithmic performance of the proposed model is compared with different other deep-learning models such as Feedforward Neural Network (FFNN), 1D Convolutional Neural Network (1D CNN), Bidirectional Long Short-Term Memory (Bi LSTM), and Long Short-Term Memory (LSTM), and a basic hybrid WT-LSTM model. All these deep-learning models are tested on two datasets of distinct sites i.e. 1A DKASC, and 1B DKASC Alice Springs Solar-PV system, from 01 January 2019 to 31 December 2019 with 5-minute resolution. The simulation results show a significant improvement in all the performance metrics obtained by the proposed improved WT-LSTM model compared to other competing deep-learning models. Specifically, the best performance values for a 15-minute forecasting horizon are: MAE (Mean Absolute Error) of 0.38617 and 0.29736, RMSE (Root Mean Square Error) of 0.78038 and 0.64868, and R² values of 0.95423 and 0.95592 across two datasets. Furthermore, MAE, RMSE, and R² values are also favorable for the 30-minute and 60-minute forecasting horizons, indicating that the proposed improved WT-LSTM model delivers superior forecasting accuracy. As these simulation results indicate that the proposed hybrid deep learning method increases the prediction precision significantly, it is useful for generation planning and reserve estimation in renewable-dominated power systems.