Short-Term Photovoltaic Power Prediction Based on LGBM-XGBoost

Abstract

Improving the accuracy of short-term power prediction of photovoltaic power generation systems is of great significance for the safe scheduling and stable operation of power systems. The algorithm in this paper uses the Light Gradient Boosting Machine (LGBM) algorithm in the baseline layer and the Extreme Gradient Boosting (XGBoost) algorithm in the boost layer, thus proposing a double-layer combined prediction algorithm based on LGBM-XGBoost to predict short-term photovoltaic power generation. This paper selects the power data of a photovoltaic power plant for one year as the training set, and predicts the actual power of photovoltaic power in the next week in the test set. Firstly, the importance of the features provided in the training set data is sorted, and the most important feature is discovered. However, the test set does not have this feature. Therefore, the strategy of divisional training is used to predict the most important feature first, and then predict the final actual power. At the same time, in order to avoid the defect of using a single prediction method for two trainings, a double-layer combined prediction method is adopted. The baseline layer uses the LGBM algorithm to predict the missing strong features of the test set. After adding the predicted strong features, establish multiple time zones for feature cross processing to build new features. Based on feature engineering, the boost layer uses the XGBoost algorithm to predict the photovoltaic power. Compared with using a single algorithm, the double-layer combined prediction algorithm effectively improves the accuracy of short-term prediction of photovoltaic power and meets the requirements of short-term prediction of photovoltaic power systems.