Advanced LSTM-GAN Model with Attention Mechanism for Carbon Footprint Assessment and Low-Carbon Strategy in Photovoltaic Energy

Abstract

Accurate power load predictions are crucial for optimizing energy management and reducing carbon emissions in photovoltaic power generation. This study presents an advanced prediction model that combines Long Short-Term Memory (LSTM), Generative Adversarial Networks (GAN), and Attention Mechanisms to effectively forecast power loads, assess carbon footprints, and develop emission reduction strategies for photovoltaic enterprises. By incorporating the Gated Recurrent Unit (GRU) module, this study trained and evaluated the model across multiple datasets, demonstrating superior performance with over 95% accuracy, 93% recall, 92% F1 score, and 92% AUC metrics, outperforming other methods. The model also showed enhanced efficiency with fewer parameters, reduced floating-point operations, shorter inference times, and lower training times. On the Sun Dance dataset, it achieved a 46.2% reduction in parameters, 48.4% in FLOPs, 49.7% in inference time, and 37.6% in training time compared to existing methods. Ablation experiments confirmed high accuracy with Mean Absolute Error (MAE) below 19, Mean Absolute Percentage Error (MAPE) below 6%, Root Mean Square Error (RMSE) below 4, and Mean Squared Error (MSE) below 6. This research not only addresses time series prediction challenges but also offers potential applications across various fields, with future work aimed at improving dataset adaptability and model interpretability.