An interpretable generalized additive neural networks for electricity theft detection in smart cities using balanced data and intelligent grid management

The growing demand for electricity calls for efficient energy use in Smart Cities (SC). Balancing imbalanced data enhances model accuracy but may complicate training and obscure true patterns in energy consumption. In this manuscript, an interpretable generalized additive neural network for electricity theft detection in smart cities using balanced data and intelligent grid management (IGANN-ETD-SCBD-IGM) is proposed. Initially, input data are collected from the Smart Grid Corporation of China (SGCC) dataset. The Reduced Noise-Synthetic Minority Over-sampling Technique (RN-SMOTE), which creates synthetic samples for the minority class and decreases noise in the majority class, is then used to balance the unbalanced data. The balanced data undergoes pre-processing using the Reverse Lognormal Kalman Filter (RLKF) to address missing values, handle outliers, and normalize the data. Then the pre-processed data are fed to Multi-dimensional Spectral Graph Wavelet Transform (MSGVT) for extracting the statistical features like Entropy, Mean, Contrast, Kurtosis, Correlation and Variance. Then extracted features are fed to Interpretable Generalized Additive Neural Networks (IGANN) which is used for detecting and classifying electrical thefts such as electrical criminals and real consumers. In general, IGANN does not express adapting optimisation techniques to identify ideal parameters to guarantee accurate electrical theft detection. The Secretary Bird Optimisation Algorithm (SBOA) is used to optimise the IGANN weight parameter. Then the proposed IGANN-ETD-SCBD-IGM is excluded in Python and the performance metrics like Precision, Accuracy, Recall, F1-Score, False Positive Rate (FPR), False Negative Rate (FNR), Matthews Correlation Coefficient (MCC) and Rate of Curve (ROC) are analysed. Performance of the IGANN-ETD-SCBD-IGM approach attains 99% higher accuracy, 94.12% higher precision and 91.92% higher recall when analyzed through existing techniques like exploiting machine learning to tackle peculiar consumption of electricity in power grids: A step towards building green smart cities (TPC-EPG-GSC-KNN), a hybrid evolutionary and machine learning approach for smart city planning: digital twin approach (SCP-DTA-SVM) unbalanced data handling techniques for classifying energy theft and defective meters in the provincial electricity authority of Thailand (UDH-CET-PEA-CNN), methods respectively.