Load Forecasting for Virtual Power Plants Based on Probabilistic Optimization and Improved LSTNet

To address the issues of insufficient prediction accuracy caused by the strong randomness and complex patterns of virtual power plant loads, this paper proposes a virtual power plant load forecasting method integrating probabilistic optimization and an improved LSTNet, aiming to enhance the accuracy and robustness of load forecasting. The method first identifies typical load operation patterns via the K-means clustering algorithm, and then introduces the Gaussian Mixture Model (GMM) to perform probabilistic optimization on the clustering results, which effectively reduces errors induced by subjective classification and improves the precision of load pattern division. At the model level, the traditional LSTNet is modified by incorporating an encoder-decoder structure, and combining Convolutional Neural Networks (CNN) with the Transformer self-attention mechanism. This enables the model to extract both local fluctuation features and capture global time-series dependencies of load data. Experimental results demonstrate that, compared with conventional forecasting methods, the proposed model achieves significant improvements in core evaluation metrics including RMSE, MAE and MAPE. The predicted curve fits the actual load variation trend better, featuring stronger generalization ability and stability. It can better adapt to virtual power plant operation scenarios with complex load patterns and multi-source coupling, providing reliable support for the efficient regulation and optimal energy allocation of virtual power plants.