In recent years, with the booming development of computer technology, the application of deep learning methods in power load forecasting is of great significance to improve the forecasting accuracy. To this end, the article uses stacked self-encoder and entity embedding to improve the multilayer perceptron, and constructs a power load forecasting model based on the multilayer perceptron. The effectiveness of the model proposed in this paper is verified on relevant datasets and examples, and a power scheduling decision-making method is proposed and optimized for scheduling. The experimental results show that the various anomalies existing in the original data processed using the model in this paper are basically effectively corrected. After the regulation strategy and optimization adjustment, the original load curve occurred peak elimination and valley filling behavior. This proves the effectiveness of the model and adjustable capacity optimization method proposed in this paper.