Variational Autoencoder-based High-dimensional Feature Extraction for Economic Analysis of Power Cost Data

Abstract

The purpose of this paper is to extract the characteristics of power cost data using a deep learning model and to evaluate and predict the cost structure, profitability, and future development trends of power enterprises by combining economic analysis methods. Firstly, the paper innovatively employs a Variational Autoencoder for feature extraction. This model extracts low-dimensional latent representations of the data through an encoder and reconstructs the data through a decoder, retaining key structural information. The dataset used here consists of 1,800 records, which include costs, revenues, output, and energy consumption data from power companies, covering multiple enterprises and time periods. Secondly, during the model training process, optimization is performed with a learning rate of 0.001, a batch size of 64, and 50 training epochs. Performance comparisons under different hyperparameter combinations indicate that the model with 256 hidden nodes in both the encoder and decoder layers yields the best performance. Lastly, economic analysis methods, such as cost-benefit analysis and economic forecasting, are applied to assess and predict the profitability and future trends of the power companies. The specific results show that the reconstruction error of this model is 0.032, and the KL divergence is 0.006. In terms of refined economic analysis, the net profit predicted by the model reaches 5.36 million yuan, with a prediction accuracy of 93.5%. In terms of robustness, although the prediction accuracy fluctuates slightly, it remains high overall, and both the training time and prediction time show stability. Moreover, testing on multiple datasets from sources such as University of California, Irvine, Kaggle, and government open data platforms shows that the model's prediction accuracy remains between 92.3% and 94.2%, with stable training and prediction times, demonstrating its strong generalization ability. The proposed model offers several advantages. Overall, this paper presents a novel approach and method for economic analysis and decision-making in power enterprises, which holds significant practical value.