Machine Learning Framework Integrating Features Selection and Gauss-Markov Linearity Assumptions for Identifying Appropriate Energy Theft Detection Modelling

Machine Learning in Energy Theft Detection (ETD) predominantly lacks studies on data features and attribute identification as the precursor to selecting the proper machine learning (ML) algorithm, thus compromising its performance. The Advanced Metering Infrastructure (AMI) provides abundant fine-grain energy data, a heteroscedastical attribute that significantly challenges the linearity model with overfitting and misclassification [1]. Since machine learning model and data are interdependent, it becomes non-trivial in optimising the ML performance when the machine learning hyperparameters and data are intertwined, influencing the suboptimal performance result. For this reason, this study's objective is to identify significant data features and appropriate ML algorithms for ETD modelling and to answer the question of the significant data features and machine learning models for ETD modelling. In achieving this, this paper establishes a framework that studies highly dense home energy data through the Gauss Markov Linearity Assumptions (GMLA) and performs a map fit gap analysis (MFGA) as the evaluation scheme to confirm data features, target selection, and the machine learning model for ETD. The outcome of the GMLA analysis has significant statistical evidence to support that the home energy data features are heteroscedastic and non-linear, as against the presumption that ohm-laws dictate the linearity system governs the energy data. Thus, the Voltage (V), Phase Angle (D), Power Factor (F) and timestamp (T) are the selected features for ML in ETD. The MFGA scheme also reveals the XGBoost gbtree and its regularisation hyperparameters as the effective ML model in ETD.