High-resolution PV power prediction model based on the deep learning and attention mechanism

Photovoltaic (PV) power generation is associated with volatility and randomness due to susceptibility to meteorological parameters intermittency. This poses difficulty in achieving the desired accuracy of PV power prediction with traditional models. Thus, this paper proposes a new predictive model based on deep learning techniques, optimized by the Bayesian optimization algorithm, to forecast a day-ahead PV power generation in high-resolution time steps. A systematic algorithm is introduced to improve time-series data quality via identifying missing samples in high-frequency datasets and imputing the missing values through the LASSO regression technique. The two data transformers for time and wind features are proposed to enhance their contributions, while other weather information, such as temperature and humidity, are considered. The proposed hybrid model incorporates CNN and BiLSTM to learn spatial and temporal patterns; moreover, the attention mechanism determines the weight values for input series and puts explicit attention on more essential parts to improve accuracy. Finally, the performance of the proposed model is compared with nine deep learning models, which are all optimized by the Bayesian optimization technique. The prediction performance comparison on actual data for a year reveals the superiority of the proposed model with the overall performance of 0,247, 0,232, 1,58%, and 0,461 in MAE, MSE, MAPE, and RMSE, respectively.& COPY; 2023 Elsevier Ltd. All rights reserved.

Automated summary

This paper proposes a new predictive model based on deep learning techniques and the Bayesian optimization algorithm for day-ahead PV power generation prediction in high-resolution time steps. The model improves time-series data quality by identifying missing samples in high-frequency datasets and imputing the missing values through the LASSO regression technique. It incorporates CNN and BiLSTM to learn spatial and temporal patterns and uses an attention mechanism to improve accuracy.