Research on a novel distribution network line loss prediction model based on spatiotemporal feature fusion neural network

Abstract

Massive penetration of distributed generation and drastic shifts in end-user electricity consumption profiles have created formidable obstacles to precise line loss monitoring within contemporary distribution networks. Targeting the problem of high-accuracy distribution network line loss forecasting, this paper develops a novel data-driven machine learning framework. Taking nodal power flow operating conditions of distribution grids as model inputs, the proposed framework first adopts the Minimum Redundancy Maximum Relevance (mRMR) algorithm to conduct discriminative feature screening. Convolutional Neural Networks are utilized to extract spatial correlations and topological inherent patterns across the distribution network, while Long Short-Term Memory (LSTM) layers characterize temporal evolutionary trends and long-range state dependencies of grid operation. Bayesian Optimization (BO) is further embedded to implement automated hyperparameter calibration, which substantially elevates the model’s predictive accuracy and generalization capability. Comprehensive case studies are carried out on standard IEEE distribution test feeders. Comparative benchmarks against mainstream state-of-the-art machine learning approaches verify the superior efficacy and prediction precision of the presented spatiotemporal feature fusion neural network.