Research on Power System Load Forecasting Based on Physical Constraints and Multi-Model Bayesian Fusion

Nowadays, the utilization of distributed power sources and renewable energy is increasing. At the same time, it is quite difficult to grasp the complex patterns of electricity consumption changes, which brings a lot of challenges to the prediction of power load. In this paper, we propose a new prediction method, which is called the Bayesian fusion prediction method. The fusion of this method is reflected in two aspects. One is the ability to extract the characteristics of electricity consumption data from multiple aspects, and the other is the consideration of actual physical constraints. First, we need to construct a hybrid structure composed of a convolutional neural network and a long short-term memory network, which also includes a one-dimensional convolutional layer and a long short-term memory layer. The one-dimensional convolutional layer is used to identify the change characteristics of electricity consumption data in a short period, such as the sudden changes in electricity load or the shape of the electricity consumption peak. The long short-term memory layer can discover the patterns of electricity consumption data over a longer period. In this way, we can understand the characteristics of electricity consumption data from different time scales. Next, through a strategy called Bayesian random training, and combined with a loss function that takes physical constraints into account, we improve the accuracy of this prediction system. Finally, we also propose a dynamic fusion framework. This framework can measure the uncertainty of each parameter in the model and use the changes in the attention mechanism of the Transformer technology to provide a probability range for the prediction. The experimental results show that compared with the commonly used prediction models in the past, the fusion network model proposed in this paper has great advantages in terms of prediction accuracy.