(ICIOT 2023) Distributed Photovoltaic Power Generation Prediction Based on Feature Extraction and Multi-Model Fusion

With the improvement of photovoltaic grid-connected power generation and the accelerated development of distributed photovoltaics, distributed photovoltaic power generation prediction plays an important role in guaranteeing the safety and stability of power grid operation. Accurate distributed photovoltaic power generation prediction is highly important, and distributed photovoltaic power generation is affected by the combination of several meteorological variables, so the deep feature extraction of meteorological variables is very critical. This paper proposes a meteorological feature extraction method for distributed photovoltaic power generation prediction and a photovoltaic power generation prediction model based on residual connection fusion of multiple models. The feature extraction method enriches the model's input by performing deep feature extraction on time, meteorological, and power generation data using statistical methods, feature cross, periodic information, approximate entropy, and photovoltaic panel temperature feature extraction methods. In the model construction, a multi-model fusion method based on residual connection is established. Firstly, a softmax regression prediction model based on KNN is proposed. Then, the overall structure of the model integrates similar day regression prediction models, TabNet, XGBoost, RandomForest, and LightGBM models, and through residual connection and multi-layer stacking, the accuracy of photovoltaic power generation prediction is continuously improved. Experimental results show that the feature extraction method and the model pro-posed in this paper are superior to other models, and can effectively improve the accuracy and stability of distributed photovoltaic power generation prediction.