A Hybrid CNN–LSTM Framework for Accurate Load Forecasting in Virtual Power Plants

Accurate load forecasting plays a critical role in ensuring the secure operation and efficient energy scheduling of virtual power plants (VPPs). However, the increasing integration of renewable generation, energy storage devices, and flexible demand resources brings substantial uncertainties and complex nonlinear characteristics to VPP load profiles. Conventional forecasting methods often face challenges in simultaneously learning short-term variations and capturing long-term dependencies in dynamic load sequences. To overcome these limitations, this paper develops a hybrid convolutional neural network and long short-term memory (CNN–LSTM) based forecasting framework for VPP load prediction. Specifically, the one-dimensional convolutional neural network is employed to extract hierarchical local temporal features from historical load data, while the LSTM module is designed to characterize long-term temporal correlations by exploiting its memory and gating mechanisms. A unified end-to-end training strategy is further introduced to optimize the feature extraction and sequence learning processes jointly. Extensive experiments are performed from multiple perspectives, including comparisons with state-of-the-art forecasting approaches, performance evaluations under diverse load fluctuation scenarios, robustness assessments with varying prediction horizons and input disturbances, and component-level ablation analyses. The experimental results verify that the proposed CNN–LSTM framework achieves higher prediction accuracy, stronger robustness, and better adaptability than existing methods. This study provides an effective and reliable forecasting solution for VPPs operating under uncertain and highly dynamic energy environments.