Data-complexity-based selection of electricity-loss forecasting models in 0.4 kV distribution networks

Abstract

Relevance: the growing volume of AMI and smart-meter data makes model selection a separate methodological problem in electricity-loss forecasting. A single forecasting method does not provide uniform accuracy across time series that differ in sample size, variability, seasonality, and multivariate structure.

Aim: to determine rational applicability ranges of Regression, GMDH, ARIMA, MLP and DBSCAN-MLP forecasting models using an integral criterion that combines data volume and relative variability.

Methods: the study used 912 real operational observations collected in 2021-2023, correlation analysis, Fisher transformation and Z-testing, polynomial regression, ARIMA, GMDH, MLP, DBSCAN-MLP, and MAPE-based error assessment. Data complexity was characterized by .

Results: experiments on 20 univariate and 20 multivariate datasets yielded empirical model-selection boundaries. Overall MAPE values were 5.01, 3.15, 5.05, 2.58 and 2.10% for Regression, ARIMA, GMDH, MLP and DBSCAN-MLP, respectively. The hybrid-model threshold was θ≥4.2 for univariate and θ≥3.9 for multivariate data.