Electricity theft detection (ETD) is a serious issue and needs proper attention to investigate and detect the fraudulent consumers. Fraudulent consumers steal electricity and burdensome utility providers, which causes huge revenue losses. To minimize such revenue loss an affine investigation and detection scheme is required. In this paper, we use AttenLSTMInception in order to investigate the electricity theft scenario. Initially, benign consumers data is manipulated through false data injection techniques (FDIs). Six various types of FDIs are used to manipulate the benign class data, which synthesizes six variants of the manipulated data. The benign and synthesized variants are concatenated and the data is fed to borderline SMOTESVM. In order to tackle the issue of synthesizing data from overlapped samples, borderline SMOTESVM is used. The balanced data is segmented into training and testing data. Furthermore, Matthews correlation coefficient is used to investigate model's performance by measuring confidence score of the binary classification. Moreover, to evaluate the robustness and to validate the performance and of the model, testing data of 20%, 40% and unseen is used. Similarly, six activation functions are used to evaluate the model's robustness. Furthermore, impact of the imbalanced data is studied using MCC score. Support vector machine (SVM), random forest (RF), decision tree (DT), convolutional neural network-based long short-term memorization (CNN-LSTM) are used as base models. Our model outperforms the base models by achieving high recall, precision, accuracy, AUC-score and F1 score. These values for our proposed model are 0.94%, 0.94%, 0.98%, 0.98%, respectively.