Performance comparison with industry benchmarks.

With the development of smart grids and the complexity of power marketing, accurately predicting users behavior and recommending suitable products to users become more and more important in power marketing. However, current methods still have some problems such as data sparsity, cold-start problem, fixed recommendation strategy and hard to adapt users dynamic behavior. It affects the recommendation accuracy of power marketing and bring bad experience to customers. To solve these problems, we propose a new neural network model for user behavior prediction and personalized recommendation in power marketing. Our model uses Graph Convolutional Networks to model user-product interaction relationship, Deep Deterministic Policy Gradient to optimize recommendation strategy in dynamic ways, and Multi-Layer Perceptron to predict user behavior. These three models work together and use their advantages to improve recommendation accuracy, adaptability and user experience. Our experiments show that compared with traditional methods, our model improves recommendation precision, recall and user-related metrics significantly. Specifically, compared with state-of-the-art (SOTA) methods, our model achieves an average improvement of approximately 5.4% in Precision, 7.7% in Recall, and 3.3% in AUC. The GCN, DDPG and MLP enhance the model‘s ability to handle multi-dimensional user behaviors and adapt to user‘s real-time feedback. This work uses neural network model to predict user behavior and recommend in power marketing in more accurate, personalized and dynamic ways. It brings better customer experience and improve business efficiency.