Research on Short-term Photovoltaic Power Prediction Considering Meteorological Fluctuations

Background: The short-term prediction of the power of photovoltaic power generation systems can promote the consumption of the power grid. Due to the influence of climate conditions, the photovoltaic power generation system has a certain volatility, however, the existing generation power prediction models require a large amount of data to serve the model training, which might reduce the reliability of the prediction results. Meanwhile, the published test models often do not describe the network configuration or the author's reasoning method of selecting parameters for the network during the training process in detail, making it difficult for further studies. Objective: The scientific aim of the work is to apply the new model to a case study photovoltaic (PV) system with two input scenarios: accurate historical meteorological information and projected meteorological data. The research aimed to develop a model designed for short-term solar energy output projections, using the Long Short-Term Memory (LSTM) network algorithm, while accounting for fluctuations in meteorological conditions. This study extends the existing research by evaluating the performance of the LSTM-based model through an analysis of prediction accuracy and error rates and comparing them with various LSTM configurations and conventional timeseries forecasting techniques within the test dataset. Methods: The architecture of the model, grounded in the LSTM algorithm, has been customized to align with the distinct meteorological trends and operational parameters of a region situated in Southwest China. The performance of the LSTM-based model is assessed by analyzing the prediction accuracy and error rates against those of various LSTM configurations and conventional time-series forecasting techniques within the test dataset. The most suitable hyperparameters of the LSTM model are finally selected. Finally, the proposed model is applied to a case study photovoltaic (PV) system with two input scenarios: accurate historical meteorological information and projected meteorological data. Results: The optimal parameters for the LSTM model in forecasting PV power are four hidden layers with 100 nodes. The comparative analysis of the outcomes reveals the average absolute percentage error is significantly higher in the second scenario compared to the first (10.314% versus 3.316%). To address the inaccuracies introduced by the weather forecast data, new features are proposed to enhance the accuracy of the LSTM model. Applying this prediction model to the case system reduces the PV power error ranging from 10.314% to 9.32%. Discussion: The optimal parameters for the LSTM model in forecasting PV power are four hidden layers with 100 nodes. Integrating forecast meteorological data into the training process enhances the model's ability to predict meteorological conditions. Conclusion: During the establishment of the prediction model, the consideration of meteorological fluctuations and the selection of appropriate models are of beneficial to the improvement of the predicting accuracy of the photovoltaic power. Moreover, this study not only provides a direction for selecting suitable characteristics and structure of LSTM model, offering an effective method for short-term output prediction of photovoltaic power generation systems.