A measurement error prediction framework for smart meters under extreme natural environment stresses

•The WPCA method is used to convert several key environmental stresses into an environmental composite index. Future inputs to the measurement error prediction model can be provided based on the characteristics of the environmental index.•A hybrid selection operator based on the optimal individual retention strategy and the sine selection probability is proposed to improve the convergence of genetic algorithm. A multiple adaptive crossover operator that can adjust the crossover probability and a multiple adaptive mutation operator that can adjust the mutation probability and step size are presented to improve the global optimization performance of genetic algorithm.•A measurement error prediction framework for smart meters is given by combining MAGA-BPNN and ARMA, which has excellent prediction performance and generalization ability. Accurate prediction of measurement errors is essential for the operational status assessment of smart meters. So far, however, there has been little discussion on the prediction of performance metrics for instruments under extreme natural environmental stress. So, we propose a new measurement error prediction framework in the paper. First, the Weighted Principal Component Analysis (WPCA) is used to convert environmental factors with correlations into an environmental comprehensive index to simplify the structure of our prediction model. Second, a new Multiple Adaptive Genetic Algorithm (MAGA) is proposed based on individual fitness and iteration times to optimize the initial weights and thresholds of the Back Propagation Neural Network (BPNN). A new measurement error prediction model, namely MAGA-BPNN, is then given based on MAGA and BPNN. Besides, the prediction residuals of MAGA-BPNN are corrected using Auto-Regressive Moving Average (ARMA). Finally, the prediction framework composed of MAGA-BPNN and ARMA is experimentally validated using real-world data from a high-dry-heat area. The mean absolute error of predictions is 3.80e-3, and the mean square error is 3.29e-5. The results show that our framework has excellent prediction ability and generalization ability. It can provide a basis for device maintenance and replacement, especially under extreme natural environmental stresses.