Abstract
This study presents a novel feature-engineered-natural gradient descent ensemble-boosting (NGBoost) machine-learning framework for detecting fraud in power consumption data. The proposed framework was sequentially executed in three stages: data pre-processing, feature engineering, and model evaluation. It utilized the random forest algorithm-based imputation technique initially to impute the missing data entries in the acquired smart meter dataset. In the second phase, the majority weighted minority oversampling technique (MWMOTE) algorithm was used to avoid an unequal distribution of data samples among different classes. The time-series feature-extraction library and whale optimization algorithm were utilized to extract and select the most relevant features from the kWh reading of consumers. Once the most relevant features were acquired, the model training and testing process was initiated by using the NGBoost algorithm to classify the consumers into two distinct categories ("Healthy" and "Theft"). Finally, each input feature's impact (positive or negative) in predicting the target variable was recognized with the tree SHAP additive-explanations algorithm. The proposed framework achieved an accuracy of 93%, recall of 91%, and precision of 95%, which was greater than all the competing models, and thus validated its efficacy and significance in the studied field of research.