Power Grid Information Prices Prediction Based on Locally Linear Reconstruction

Information prices serve as critical references for cost management in equipment and materials, which constitute a substantial portion of power grid construction projects. Typically, information prices for a given period are estimated based on procurement and bid prices collected during the preceding period. However, procurement and bid data are often sparse, covering only a subset of the total equipment and materials. The large number of items further complicates the process, making manual price querying time-intensive, while the lack of sufficient historical data limits the development of robust models. Principal Component Analysis (PCA) reveals a high degree of collinearity within the data, prompting the development of a Locally Linear Reconstruction (LLR) algorithm for price prediction. LLR predicts prices by preserving the locally linear structure of the data. For each item, its k-nearest neighbors are identified based on a combination of name similarity—calculated as the cosine similarity of Term FrequencyInverse Document Frequency (TF-IDF) representations—and price correlation, determined using the Pearson Correlation Coefficient. Prices for each item are then predicted using Least Absolute Shrinkage and Selection Operator (LASSO) regression applied to the prices of its identified neighbors. Empirical results demonstrate that LLR achieves a low Mean Absolute Percentage Error (MAPE) and effectively identifies potential outliers. The algorithm introduces greater independence in prediction outcomes, making it particularly suitable for integration with other models through ensemble learning. Furthermore, LLR obviates the need for extensive feature engineering, reducing the time required for large-scale price prediction tasks.